TL;DR

• We tested cold winters theory using polygenic scores from ancient genomes matched to historical winter temperatures dating back 21,000 years.

• Height and educational attainment polygenic scores significantly correlate with colder winters (p < 0.001), even after controlling for ancestry, in both modern and ancient samples.

• Cognitive ability and non-cognitive ability polygenic scores show no significant correlation with winter temperatures after ancestry controls.

• Cold winters theory receives partial support. The correlations exist but are weaker than proponents might expect.

A few weeks ago I was chatting with Davide Piffer, and we arrived at the topic of ancient genomes and cold winters.

Cold winters, proposed by Richard Lynn, is an old theory that individuals subjected to hard cold winters were evolutionarily selected for long-term planning and intelligence; those who didn’t prepare died. It’s easy to imagine how adaptation for these traits would then give rise to agriculture or future advancements along the human evolution and tech tree to the civilization we have today.

Piffer told me that, controlling for ancestry, there was no correlation between latitude and an ancient genome’s polygenic score for cognitive ability. Height polygenic scores survived the controls, which made sense: taller, larger bodies have smaller surface areas relative to their volume, preserving heat. This is a classic ecogeographic pattern (Bergmann’s rule). But the cognitive measures didn’t show the same robustness.

Not long after, he published a substack article on the topic. I then tweeted his results which went viral relative to my previous posts.

Summary of Piffer’s Post

Piffer used polygenic scores for educational attainment (EA), cognitive ability (Cog), non-cognitive ability (NonCog), and height, tested against absolute latitude as a proxy for climate. He controlled for the first 10 genome-wide principal components (ancestry), genomic coverage (data quality), and time (years before present). He ran analyses on ancient samples, modern samples, and pooled data, at both individual and group levels.

His findings: in naive models without ancestry controls, all traits showed strong latitude gradients. After PC adjustment, EA, Cog, and NonCog lost their latitude signal. Height alone retained a significant positive association with latitude across all samples (partial r = 0.05-0.06, p < 10^-10).

Piffer was explicit about the limitations of his approach. He acknowledged that PC adjustment is inherently conservative: it asks whether there’s a latitude effect that survives conditioning on genome-wide ancestry. If climate shaped ancestry distributions over deep time rather than exerting parallel selection within ancestries, PCs would absorb much of that signal by design. He framed his results as “evidence against a strong, ancestry-independent, global latitude gradient, not as a claim that climate played no role at all.”

He concluded by calling for more direct tests: “Latitude is a blunt proxy, and the theory’s core prediction concerns winter severity and seasonality over many generations, not latitude per se. More direct tests would require explicit measures of winter climate, seasonal volatility, and within-continent comparisons where environmental gradients can be evaluated without being dominated by deep ancestral structure.”

This article is that test. Piffer and I teamed up to run a more direct analysis, using actual winter temperatures rather than latitude, and applying a different set of controls and methods. Working together let us move faster than we would have separately. We shared datasets old and new, retested methods more systematically, and bounced ideas off each other. This post and his substack article are the result.

From Latitude to 21,000 Years of Winter

I needed historical temperature data going back far enough to match the ancient genomes. I found CHELSA-TRACE21k, a dataset of geospatial temperature estimates at century-level resolution dating back 21,000 years.1 Below are daily minimum average temperatures for January, 10,000 before present.

For each genome, we calculated winter temperature by averaging temperatures within a 25 km radius of its location, excluding bodies of water. We used January for Northern Hemisphere samples and July for Southern Hemisphere samples, then matched each genome to the nearest century in the temperature data (a sample 8,475 years before present uses data from 8,500 years before present).2

The correlation between absolute latitude and winter temperature was r = -0.792. Strong, but not perfect. If latitude were a perfect proxy for winter severity, switching to actual temperatures wouldn’t matter. The imperfect correlation means there’s information in the temperature data that latitude alone doesn’t capture.

A multicollinearity check confirms all three climate variables can be used in the same model (VIF = 1 / (1 - r²), where r² is from regressing each predictor on the others):

All VIFs below 5, so multicollinearity is acceptable.

The ancient genomes come from an amalgamation of sources. Piffer provided the data along with four polygenic scores (EA, height, cognitive ability, and non-cognitive ability), principal components up to PC20, and latitude/longitude coordinates for each sample.

Controlling for Ancestry

The previous analysis controlled for ancestry using principal components (PCs). The first PC is the variable that explains the most variation in the genomes, the second explains the second most, and so on. The first two PCs are often plotted to visualize population structure. Since PCs are mutually uncorrelated by definition, they provide a clean “map” of ancestry. Here’s Europe:

Controlling for too many PCs risks absorbing not just ancestry, but the very trait variation we’re trying to detect. At that point, any correlation would be rendered non-significant by construction. Additionally, more controls crowd out the regression: each added variable, even if meaningless, increases noise in the other estimates.

So, the question is: what’s the minimum number of PCs that adequately controls for ancestry?

We ran two tests to find this minimum.

First, at what point do additional PCs face diminishing returns for predicting polygenic scores? If additional PCs continue to improve prediction accuracy, ancestry retains explanatory power. Once predictions plateau, we’ve captured the ancestry component, and further PCs risk absorbing trait variance.

Second, at what point do additional PCs face diminishing returns for predicting a genome’s latitude and longitude? Latitude is our proxy for cold winters, so this tells us when geographic ancestry has been adequately controlled. Beyond that point, additional PCs no longer capture geographic structure.

Both tests point to the same answer: 3 to 6 PCs. The correct number likely lies at the intersection of these two constraints: enough to capture geographic ancestry, but before we start absorbing trait variance. We settled on 6 to be generous.

Results

We controlled for ancestry (6 PCs) and date using piecewise regression, since the relationship between years before present and EA is non-linear.

To check robustness, we tested how results changed as we added more PCs. EA and height maintained significance up to PC9.

Using 6 PCs, here are the results:

Height and educational attainment show significant effects; cognitive and non-cognitive ability do not.

The betas are in standard deviation units per degree. To put this in perspective, consider a 30°C shift (roughly the difference between Mediterranean and Scandinavian winters). For height, a 30°C decrease in winter temperature is associated with a 0.46 SD increase in height PGS. For educational attainment, a 30°C decrease is associated with a 0.14 SD increase. The winter betas are negative because colder temperatures (lower values) predict higher polygenic scores.

We also ran a regression with winter temperature, summer temperature, and latitude together. As discussed earlier, multicollinearity between these variables is low (VIF < 3), so the betas remain stable. When correlated predictors compete in the same model, the one capturing the true causal signal typically retains or increases its effect size, while redundant predictors attenuate.

For EA, colder winters and warmer summers independently predict higher polygenic scores. For height, latitude dominates: higher latitudes predict taller stature, with cooler summers contributing independently.

The table results come from a regression predicting each trait from the climate variable, controlling for ancestry (6 PCs) and date. To visualize these relationships, we removed the effect of the controls from the trait, then plotted what remains against each predictor. This isolates the predictor’s effect but may look slightly different from the table betas, which estimate all predictors simultaneously.

A different way to visualize this: plotting the controlled height PGS onto a map of Europe. This is after removing the effect of ancestry and date.

The dataset also includes 3,111 modern samples (date BP = 0). Running the same regression on moderns corroborates the main findings: for EA, winter temperature remains the dominant predictor, while summer and latitude lose significance. For height, all three predictors remain significant, with latitude showing the strongest effect.

But These Correlations are Tiny!

They are small, but multiple factors suppress them.

First, linkage disequilibrium. Polygenic scores trained on one ancestry group become less accurate when applied to others, not biased in any direction, just noisier. The ancient samples span the globe and date back tens of thousands of years, which increases measurement error and attenuates correlations.

Second, polygenic scores aren’t that predictive to begin with. The height PGS accounts for 40-45% of phenotypic variance in European ancestry populations. The EA PGS (EA3, EA4) accounts for only 12-16%. Since height’s polygenic score captures far more of the trait’s variance, we’d expect stronger and more detectable effects for height, which is exactly what we observe.

Despite all this, the correlations survived and remained statistically significant.

Is cold winters theory dead? I don’t think the answer is a simple yes or no. The correlations aren’t what proponents might have expected, but they’re not zero either. Detractors can’t claim the theory is meaningless.

• AI capabilities are accelerating. Aggregate benchmark performance shows the rate of progress roughly doubling since 2024.

• The duration of the longest task AI can autonomously complete (METR) doubles every 4 months.

• On fluid reasoning (ARC-AGI), AI went from 18% to 86% (human-level) in one year.

• On expert knowledge questions beyond most humans (Humanity’s Last Exam), AI performance increased from <5% to 40%.

• AIs have general intelligence. Performance inter-correlates across all benchmarks, just as human cognitive abilities correlate across all cognitive tests.

• AIs moved from next-word prediction to direct problem-solving. This produced emergent reasoning: models spontaneously began thinking before answering.

• Company P/E ratios are half those of the dot-com era. Across AI companies, revenue grows faster than valuation (~4x/year vs ~3x/year). Enterprise API spending doubled in six months. Total AI output across major providers via API increased 9x in 8 months to ~0.75 trillion words per week.

• AI adoption outpaced the internet. Worker usage doubled 2024-2025. US businesses with paid AI subscriptions went from 26% to 45% since January 2025.

• GPT-4 level intelligence now costs 1,000x less than three years ago, with similar deflationary trends at other capability levels.

Introduction

A new age is dawning.

AI capabilities are advancing faster than most realize. Progress on benchmarks has accelerated since 2024, driven by a fundamental shift in training methods. The economic adoption is real: billions of users, exponentially growing enterprise use, productivity gains measurable across professions. The bubble arguments don’t hold.

This article makes the case systematically.

First, we establish that AI possesses general intelligence. Using item response theory (the same statistics behind IQ tests), we compare AI’s capability structure to human cognition, finding parallels and nuance.

Second, we examine performance on key benchmarks: fluid reasoning (ARC-AGI), autonomous task completion (METR), and expert knowledge (Humanity’s Last Exam). On all three, AI has improved dramatically in the past year, with some approaching human-level performance at a fraction of the cost.

Third, we document the acceleration. Both METR’s time horizons and aggregate performance across all benchmarks show the rate of progress roughly doubling since 2024. We explain why: a training paradigm shift from next-word prediction to directly optimizing for problem-solving through reinforcement learning has produced emergent reasoning, a capability nobody programmed.

Fourth, we address the skeptics. Scaling bottlenecks exist but are surmountable. The bubble arguments fail: P/E ratios are half those of the dot-com era, OpenAI’s revenue is growing faster than its valuation, enterprise API spending is doubling every six months, and worker adoption doubled year-over-year.

If these trends continue, and no fundamental barrier has yet appeared, human-level AI across most domains arrives within this decade. The implications are difficult to overstate.

Understanding AI’s Intelligence and Progress

Epoch Capabilities Index: AI’s “IQ”

Intelligence is a concept, IQ is a test, and ‘g’ (general intelligence) is a statistical phenomenon. Performance on all cognitive tests correlates, and IQ tests attempt to extract the construct underpinning this: general intelligence, or ‘g’. AIs do not lack general intelligence.

Using item response theory, the same statistics behind IQ tests, Epoch AI published the Epoch Capabilities Index, an “IQ” for AI. The score is based on 1103 benchmarks across 126 AI models (Ho et al., 2025).

The methodology constructing ECI is similar to IQ tests for humans.1 To analyze the statistical similarities, we compared ECI’s structure to human g extracted from 19 cognitive tests in the Vietnam Experience Study.

On human cognitive tests, performance on each subtest correlates with all others. Epoch AI reports this holds for AI benchmarks:

The table below shows the weights on the different benchmarks in this component, accounting for 80% of the total weight. Note that the *weights are all positive* and not very dispersed.

The first principal component explains similar variance in both human cognitive tests (44%) and AI benchmarks (40-50%). This parallel is meaningful: AI capabilities emerged with a human-like factor structure without being explicitly designed that way.

AI’s tighter factor structure (higher subtest correlations, more variance in the top 2 PCs) could reflect more unified cognition. However, three methodological factors likely explain it:

First, AI benchmarks have hundreds of items versus human subtests with 10-30, reducing measurement error.

Second, model responses are deterministic. Cremieux found insufficient variation in Claude 3’s outputs to construct “IQ” scores between response variations. This determinism is consistent across benchmarks.

Third, AI models are derivatives of each other (Claude 3 to 3.5 to 4, GPT-3 to 4 to 4o), which explains the large variance in the second principal component. Epoch’s post on ECI shows “Claude-y” (the 2nd PC) is highest on Anthropic models and lowest on OpenAI’s. AI performance captures both general capability and company-specific training focus.

Strangely, on matrix problems Cremieux reported no correlation between the % of humans and % of Claude responses that got each question correct.

This result is suspicious. If it generalized to all cognitive items, there would be no correlation between item difficulty for AIs and humans. Taken to the extreme: is proving the Riemann Hypothesis really no harder than elementary school maths for AI?

We can test this. ECI performance across three mathematics benchmarks shows that between models, ECI correlates with mathematical outcomes.

As problem difficulty increases from elementary school mathematics (GSM8K) to frontier problems (FrontierMath), praised for their difficulty by Terence Tao, AIs are less likely to solve them. Obviously, humans follow the same pattern. One might argue AIs are just parrots regurgitating memorized answers, but these benchmarks keep questions private to prevent contamination.

But ECI itself is not interpretable. What does an ECI of 140 mean? Or an increase of 15 per year?

To make this concrete, we focus on specific benchmarks: fluid problem solving, autonomous task completion, and expert knowledge. These allow direct comparison to human performance.

Fluid Problem Solving: ARC-AGI

ARC-AGI is Raven’s Progressive Matrices for AI. It’s arguably better than RPM: not multiple choice, with many possible solutions per problem, so patterns cannot be memorized. It still measures the same core pattern matching ability.

ARC-AGI has two versions: an “easy” benchmark (ARC-AGI-1) and a “hard” benchmark (ARC-AGI-2). Example questions from each:

You can attempt both questions yourself here and here, or take the full test.

To prevent contamination, AI models are evaluated on a private dataset not accessible to the internet. The current leaderboard shows the general trend: improving performance at declining cost.

Humans perform well on ARC-AGI. The benchmark was developed to demonstrate that AIs lack true intelligence, that they’re just pattern matching from training data.

This is no longer true. AI performance has skyrocketed from 18.0% to 86.2% in one year, now matching average human performance at 3x cheaper cost than an MTurker2, or 17x cheaper with Gemini 3 Flash. Benchmark saturation is near for ARC-AGI-1.

At this rate, AI will surpass the average STEM graduate at 3-20x cheaper cost by next year.

For ARC-AGI-2, AI barely reached 5% last year. Now it’s at 52.9% at a price cheaper than a human.3

We can also use actual Raven’s Progressive Matrices from Mensa Norway. Maxim Lott runs a website where AI models attempt RPMs daily, using a withheld test to prevent contamination. This has become a popular benchmark for comparing AI to human IQ, and is analogous to ARC-AGI.

However, it should not be taken seriously:

• An IQ of 100 on this test does not correspond to 100 in the general population. The test reflects a self-selected group.

• This is not a full IQ test. Measuring ‘g’ requires multiple subtests, not just RPM. RPM has poor g-loading because patterns can be learned (Krautter et al., 2021).

• Assigning AI a human IQ is inappropriate because, as shown above, AI and human intelligence are asymmetrical.

Evo plotted frontier AI models on this Mensa Norway test. Gemini 3 Pro scores an “IQ of 130”, and performance continues to increase even on the withheld dataset.

Autonomous, Agentic Ability: METR

AI’s problem solving ability is now at human-average. But problem solving alone doesn’t mean it can integrate into the economy or complete tasks that take weeks or months.

AIs struggle with long horizon tasks for two reasons.

First, context windows. AIs are input-output machines: they take a set number of words and produce output. Context windows are limited to 100,000s to 1,000,000s of words; beyond this, AIs “forget”. The workaround is writing summaries, but AI doesn’t know what’s important long-term. Long-term memory is something AI fundamentally lacks.

Second, continual learning. Humans learn continuously from little information: we’re told what an apple is, feel it, eat it, and remember for years. AI needed most of the internet to learn what an “apple” is. The same applies to skills: a human writing a Substack learns what makes a good post over time. AI lacks this capability due to context limits and data requirements.

METR’s time horizon benchmark (Kwa et al., 2025) visualizes this. For humans, task cost scales roughly linearly with time. For AI, cost increases super-exponentially with task length.

AIs are faster and cheaper for short atomic tasks, but struggle with tasks taking a day or more.

This is changing exponentially. The improvement is so consistent it resembles a “Moore’s Law” for AI. Below are time horizons for 50% and 80% task completion success. The current state-of-the-art: Opus 4.5 at 4 hours 49 minutes (50% success), GPT-5.1-Codex-Max at 32 minutes (80% success).

To calculate time horizon for a single model: they estimate success rate by task time bin, fit a curve, then use the curve to find the time at which the model hits 50% or 80% success.

Plotting multiple models shows newer models shifting the curve rightward, achieving longer time horizons.

The doubling time across this period is roughly 6 months. A piecewise regression from 2024 onwards shows it has shortened to about 4 months. More on this acceleration later.

Like Epoch’s Capabilities Index, METR is a compilation of benchmarks. Breaking it down highlights AI’s asymmetry with human intelligence.

OSWorld, WebArena, and RLBench are visually-based agentic tasks. OSWorld (Xie et al., 2024) tests OS settings, terminal use, GUI data analysis (LibreOffice, Excel), document editing, email, and accounting. On these, AI is ~4 years behind the overall trend. We’re still in the pre-GPT-4 era for computer use via the user interface.

This highlights another limitation. Humans think and act simultaneously; our senses don’t stop when we think. AI does nothing but think when thinking: think → act → think → act. Humans can think+act in parallel.

AI cannot interact with the physical world concurrently. Using an OS with vision means taking screenshots, acting, taking more screenshots. If something unexpected happens between screenshots (a popup, an ad), it fails. This also makes AI awful at video games requiring reaction speed, like CS:GO.

Despite these limitations, progress continues. o3 (April 2025) performed 23% on OSWorld. Eight months later, Opus 4.5 performs 67.1%.

General Knowledge: Humanity’s Last Exam

Unlike agentic tasks, AI excels at general knowledge compared to humans. Trained on the internet and almost all published papers and books, even last year’s o1 scored ~80% on GPQA (PhD-level science questions). Gemini 3 Pro now scores 93%. This is superhuman: PhD experts answer 65% correctly, non-experts 34% *despite* web access.

But AIs are not omniscient. Humanity’s Last Exam (Phan et al., 2025) is the most difficult knowledge benchmark. Example questions from their website:

Consistent with other benchmarks, AI has moved from <5% to ~40% correct in one year. By end of 2026, this benchmark will likely be saturated.

AI Progress is Accelerating

The rate of frontier AI progress has nearly doubled.

This was first noticed in METR’s time horizons, where progress has accelerated since early 2024.

This is likely real. Epoch AI reports the same pattern in ECI, their aggregate capability index.

Extrapolating METR forward: if the 80% success time horizon is ~30 minutes now, by early 2027 it will be 3 hours, by 2028 18.5 hours.

A work day is 8 hours; 18.5 hours is nearly half a week. By 2028, given ARC-AGI trajectories, AI’s problem solving will likely exceed above-average humans, and it will be an oracle of expert knowledge. By 2029, given agent-friendly environments with MCPs (tools to give itself context and interact with computers in text-based formats where it flourishes), it could automate most white collar work.

Top-20 global forecaster Peter Wildeford lays out similar estimates, mirroring these calculations. Projections vary despite trend consistency; AI Futures models AI milestones under dozens of parameters, estimating AI will complete 1 work year at 80% reliability by 2030.

The acceleration is unlikely to be due to AIs assisting researchers at coding experiments. Experiments are bottlenecked by throughput (compute) and quality (research taste). Labs are bottlenecked by compute (almost all goes to R&D). Since frontier lab researchers are likely 130-145 IQ, AI must supersede them in capability, and if continual learning is not solved, must also trump cumulative human-learnt intuition on AI research.

Cause of the Acceleration: Emergent Intelligence from Reinforcement Learning

So why has progress accelerated? The answer is that there’s been a fundamental shift in how AIs are trained, from simply predicting the next word, to reinforcement learning, leading to emergent thinking.

To explain; in reinforcement learning, an agent (in the case of LLMs, an AI) tries actions in an environment like playing a game or solving a problem. The agent’s action is then associated with a reward or penalty based on the outcome; this might be losing a game (penalty) or correctly solving a problem (reward). Then, using mathematics4 the agent’s parameters are calibrated towards the configuration that resulted in the desired outcome. No labelled data is required for this, only self-play.

Reinforcement learning algorithms have previously produced superhuman performance. In 2016, DeepMind’s AlphaGo (Silver et al., 2016) defeated the world champion at Go, a game with more possible board positions than atoms in the universe. In 2022, Meta’s Cicero (FAIR et al., 2022) achieved human-level play in Diplomacy, a game requiring negotiation, deception, and long-term planning. These systems learned through self-play: millions of games against themselves, reinforcing patterns that led to victory, weakening patterns that led to defeat.

AIs are now trained in reinforcement learning environments, rather than just predicting the next word. Next word prediction produced sophisticated autocomplete; reinforcement learning produces reasoning.

Reinforcement learning caused this. Frontier models now train on verifiable tasks: for example, mathematics problems and coding exercises, where correctness can be verified with certainty. For math, the model’s answer is string-matched against ground truth. For code, output is executed against test cases in a sandbox. The model attempts problems, receives binary feedback (correct or incorrect), and adjusts its weights accordingly. This is Reinforcement Learning with Verifiable Rewards (RLVR).

The optimizer of choice is Group Relative Policy Optimization (GRPO), proposed by DeepSeek in DeepSeekMath (Shao et al., 2024). For each problem, the model generates many candidate answers. Answers scoring above the group average are reinforced; those below are penalized. The group itself provides the baseline for comparison, eliminating the need for a separate critic model. This approach has become standard; as of late 2025, open-weight frontier models have not significantly deviated from RLVR and GRPO.

The result was phenomenal.

A new property emerged with no change in architecture: reasoning. Models, for the first time, on their own, without prompting, began to reason before answering questions. Nobody programmed this. This result was first published by DeepSeek R1-Zero (DeepSeek-AI et al., 2025), which termed it the “aha moment”.

Trained with pure reinforcement learning and no supervised examples, LLMs spontaneously started generating reasoning traces before answering. After this discovery, AIs began to be explicitly trained to encourage this behavior.

It’s worth reminding the reader how these models, and neural networks (the building blocks of LLMs), work.

These models generate text autoregressively: each word produced becomes context for the next. When the model writes a reasoning step, it attends to that step and builds on it. This is not recursion or a special architecture. It is roughly the same transformer used since the pre-reasoning era, trained differently.

Couple this with what we know about neural networks. Neural nets can approximate any function given enough parameters.5 Given that LLMs are now explicitly trained and validated on outcomes requiring intelligence, this heralds a paradigm shift. It is now reasonable to assert we are creating models with general intelligence.

The timeline of accelerating AI capabilities aligns with this development; OpenAI announced o1 (the first reasoning model) in September 2024; DeepSeek released R1 in January 2025. The acceleration in capabilities coincides roughly with the frontier labs implementing this technique at scale.

AI Progress Inevitability, Bubbles and Bottlenecks

Scaling Laws and Bottlenecks

Training AIs has exponentially diminishing returns. For a given architecture, exponentially more compute is required to make the same gains. These are scaling laws, and they are universal. Even my Pedo AI saw diminishing returns in accuracy as it was fed more compute. This makes each generation of AI models exponentially more expensive to train.

To accommodate this, training compute is increasing at 4.3x per year and installed capacity at 2.3x per year. However, there is valid speculation that capacity buildout will reach limits. Since compute growth has been a main driver of AI progress, slowing buildout would slow capability gains.

GPUs are power hungry. By the 2030s, AI demand for electricity will exceed the US power grid. Epoch provides estimates for other hard limitations, such as chip production capacity. Note that data scarcity is likely not an issue: since AIs are now trained through reinforcement learning, they self-generate data. The fear that AI would run out of text to predict no longer applies.

These are not hard barriers to AI progress for the next decade:

• US energy stagnation is partially due to lacking demand. There hasn’t been economic ROI for developed countries to vastly expand their grids.

• Should the US fail, China’s electrical grid is double that of the US, and their models are only ~8 months behind the frontier.

• The cost of renewables and other energy sources is decreasing exponentially, making this bottleneck easier to solve.

• Power is a small expenditure relative to compute. AI companies could pay a premium for electricity, building their own power infrastructure.

Overall, this won’t be a massive bottleneck for the next decade, in agreement with Epoch.

However, if progress is 2-3x slower than forecasted despite capacity buildout, world-altering AI could take decades or require another architectural breakthrough. Given that progress has accelerated and how close we are to transformative AGI, this is possible but unlikely.

Bubble Woes

Compared to the Dot Com Bubble and Japanese Bubble, price-to-earnings ratios for AI companies are nearly half. From this FT article:

Revenue and valuation for AI companies are growing exponentially. For OpenAI, valuation is increasing at 3.0x per year, with revenue growing faster at 4.1x per year.

Despite this growth, parallels to previous bubbles exist. Telecommunications and railway companies were initially profitable with rapidly growing revenues. But they overinvested, supply outpaced demand, and the bubble popped. From Peter Wildeford’s substack:

Consider Britain’s Railway Mania of the 1840s. The Liverpool and Manchester Railway, opened in 1830, generated 10%+ annual returns and demonstrated railways could dramatically reduce transportation costs. This success triggered an explosive investment. Between 1844 and 1847, Parliament authorized over 8,000 miles of new rail construction.

Multiple companies laid parallel routes, each assuming they would capture market share. But a lot of the new routes were not profitable. When the crash came in 1847, thousands of investors lost fortunes. Yet the infrastructure remained valuable, powering Britain’s industrialization through the late 19th century. The technology thesis proved correct; the financial structure was catastrophic.

The telecommunications crash of 1997-2002 followed a similar pattern. The thesis was sound — explosive internet growth would require massive bandwidth capacity. Companies laid millions of miles of fiber optic cable, with industry capital expenditures reaching $600B from 1997 to 2001. But the simultaneous construction by competitors created catastrophic oversupply and a significant portion of the fiber was installed but unused. Though the fiber ultimately did end up seeing use over the next two decades, this was far too late for original investors.

Wildeford pushes back:

... unlike railway track sitting idle for years, AI data centers are being utilized immediately upon completion. OpenAI’s Abilene facility began running workloads as soon as capacity came online... The constraint is currently supply, not demand — which is why companies are aiming to build as much as possible.

AI infrastructure shows more flexibility than fiber optic cable. GPUs can run various workloads, data centers can host different services, and cloud capacity potentially retains value even if AI-specific demand disappoints.

Another bubble argument: AI expenses outweigh revenue. OpenAI won’t be profitable until 2029.

This is superficially relevant. Most spending is on infrastructure for compute, and the majority of OpenAI’s compute spending is R&D, not running models. If progress slowed, they could reduce buildout and switch to providing services. Even in a worst case, revenue streams are not threatened. This reinforces investor confidence.

Demand for the AI Economy

One might argue real AI use isn’t increasing, that it’s all hype. The opposite is true: AI demand is exploding across all sectors.

AI adoption has outpaced the internet, with billions of active users. Between 2024 and 2025, worker AI usage doubled.

Adoption extends beyond casual users to enterprise. The proportion of US businesses with paid AI subscriptions is consistently increasing: from ~26% to ~45% since the start of 2025.

Businesses are profiting. According to Stripe, AI startups generate revenue sooner post-launch than traditional software startups.

This is confirmed by the decoupling in total startup revenue between AI and non-AI startups.

Price predicts demand between models. GPT-4 level models have declined in price by 1,000x in 3 years, with similar deflationary trends at other intelligence levels. We would expect demand to increase accordingly (Demirer et al., 2025).

This is what we see. Menlo Ventures estimates enterprise API spending on generative AI increased from $3.5 billion to $8.4 billion in six months. OpenRouter reports token output increased 9x in 8 months, now at 1 trillion tokens per week (roughly 750 billion words).

This demand is driven by genuine productivity, not hype. From 100,000 conversations with Claude, Anthropic estimates timesavings of up to 80% on many tasks. This corroborates our earlier article showing AI improves job performance even with weaker models.

This has quantifiable effects on hiring. Companies are starting to prefer AI over graduates. Occupations more exposed to AI show larger headcount declines for those starting their careers (Brynjolfsson et al., 2025).

So overall, the AI economy is here, arriving faster than most impactful technologies in history. Adoption is real, productivity gains are measurable, and demand is outpacing supply.

Conclusion

AI possesses general intelligence. Using the same statistics behind IQ tests, we showed AI’s capability structure parallels human g: the first principal component explains similar variance (44% for humans, 40-50% for AI), and performance correlates across benchmarks just as it does across human cognitive tests.

On key benchmarks, progress has been dramatic. ARC-AGI went from 18% to 86% in one year, now matching average human performance at a fraction of the cost. METR’s time horizons are doubling every 4 months. Humanity’s Last Exam improved from <5% to 40%.

This acceleration stems from a paradigm shift: reinforcement learning with verifiable rewards, which produced emergent reasoning without anyone programming it. The architecture hasn’t changed since GPT-2. What changed is training AI directly on problem-solving rather than next-word prediction.

Scaling bottlenecks exist but are surmountable. The bubble arguments fail: P/E ratios are half those of dot-com, revenue is growing faster than valuation, enterprise spending is doubling every six months, and worker adoption doubled year-over-year.

My Personal Closing Thoughts

I tried avoiding making forecasts and year-by-year predictions of the future, because I want you, the reader, to make your own inferences.

However, I am of the opinion that, given the evidence, better than human-level AI across >95% of domains will arrive within 5-10 years.

Thus, in the past few short years, my expectations for the future have been completely overhauled, updating to this reality. Although AI alignment was not covered in this article, there is also concerning evidence that a world in which these capable systems work against humanity is possible. Some consider human extinction likely or even certain.

As a user since GPT-4 and an early adopter of agentic interfaces like Cursor and Claude Code, I’ve seen the time horizons expand firsthand. It could barely write a single function a few years ago, and it would be very slow at doing so. Then it could write sections of scripts and some middling logic. Then the entire script, and quickly too. Now it can interact with entire codebases, writing and executing scripts on my machine directly rather than through a browser.

This all happened in about two years.

For me, this is as real as it gets.

When I used to ponder issues humanity may face (e.g. dysgenics, population decline, etc), I could have a large degree of detachment. After all, the problems were theoretical, numbers on academic papers. Nothing I could see with my own eyes as I have done and continue to see with AI.

Furthermore, even if I could experience these other issues, its possible culmination into a civilizational collapse was lifetimes away, or just uncertain given the timescales. A lot can happen in a few decades. Humanity’s ability to innovate captures that uncertainty, as I have written about.

When speaking to normies (when appropriate) about these future issues, the response is always a variant of “that’s cool bro, so anyway”, and admittedly, that’s the correct response to large and possibly dubious claims. For me, they are also somewhat of a novelty; my telling them was driven by interest and a genuine curiosity to hear their opinion (if they were likely to have one), rather than any civic duty to “spread the word”.

But with AI, that same response hits different. I can’t help but think there’s a disconnect between what I really believe to be important and what others know.

For example, two large polls show Americans rank AI as the lowest threat and voting priority. Another survey shows only 28% even know what AI is, most think it’s a tool looking up answers in a database.

Unfortunately for the unaware, reality can only be ignored for so long, regardless of what the general population may think or feel about AI. A new age is dawning.

”Just because you do not take an interest in politics doesn’t mean politics won’t take an interest in you.”

— Pericles

References

Brynjolfsson, E., Chandar, B., & Chen, R. (2025). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Retrieved January 2, 2026, from https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/

DeepSeek-AI, Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., Zhang, X., Yu, X., Wu, Y., Wu, Z. F., Gou, Z., Shao, Z., Li, Z., Gao, Z., … Zhang, Z. (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (No. arXiv:2501.12948). arXiv. https://doi.org/10.48550/arXiv.2501.12948

Demirer, M., Fradkin, A., Tadelis, N., & Peng, S. (2025). The Emerging Market for Intelligence: Pricing, Supply, and Demand for LLMs (No. w34608; p. w34608). National Bureau of Economic Research. https://doi.org/10.3386/w34608

Ho, A., Denain, J.-S., Atanasov, D., Albanie, S., & Shah, R. (2025). A Rosetta Stone for AI Benchmarks (No. arXiv:2512.00193). arXiv. https://doi.org/10.48550/arXiv.2512.00193

James, G., Witten, D., Hastie, T., & Tibshirani, R. (2017). An introduction to statistical learning: With applications in R (Corrected at 8th printing 2017). Springer.

Krautter, K., Lehmann, J., Kleinort, E., Koch, M., Spinath, F. M., & Becker, N. (2021). Test Preparation in Figural Matrices Tests: Focus on the Difficult Rules. Frontiers in Psychology, 12, 619440. https://doi.org/10.3389/fpsyg.2021.619440

Kwa, T., West, B., Becker, J., Deng, A., Garcia, K., Hasin, M., Jawhar, S., Kinniment, M., Rush, N., Arx, S. V., Bloom, R., Broadley, T., Du, H., Goodrich, B., Jurkovic, N., Miles, L. H., Nix, S., Lin, T., Parikh, N., … Chan, L. (2025). Measuring AI Ability to Complete Long Tasks (No. arXiv:2503.14499). arXiv. https://doi.org/10.48550/arXiv.2503.14499

Meta Fundamental AI Research Diplomacy Team (FAIR)†, Bakhtin, A., Brown, N., Dinan, E., Farina, G., Flaherty, C., Fried, D., Goff, A., Gray, J., Hu, H., Jacob, A. P., Komeili, M., Konath, K., Kwon, M., Lerer, A., Lewis, M., Miller, A. H., Mitts, S., Renduchintala, A., … Zijlstra, M. (2022). Human-level play in the game of Diplomacy by combining language models with strategic reasoning. Science, 378(6624), 1067–1074. https://doi.org/10.1126/science.ade9097

Phan, L., Gatti, A., Han, Z., Li, N., Hu, J., Zhang, H., Zhang, C. B. C., Shaaban, M., Ling, J., Shi, S., Choi, M., Agrawal, A., Chopra, A., Khoja, A., Kim, R., Ren, R., Hausenloy, J., Zhang, O., Mazeika, M., … Hendrycks, D. (2025). Humanity’s Last Exam (No. arXiv:2501.14249). arXiv. https://doi.org/10.48550/arXiv.2501.14249

Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y. K., Wu, Y., & Guo, D. (2024). DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models (No. arXiv:2402.03300). arXiv. https://doi.org/10.48550/arXiv.2402.03300

Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T., Leach, M., Kavukcuoglu, K., Graepel, T., & Hassabis, D. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587), 484–489. https://doi.org/10.1038/nature16961

Xie, T., Zhang, D., Chen, J., Li, X., Zhao, S., Cao, R., Hua, T. J., Cheng, Z., Shin, D., Lei, F., Liu, Y., Xu, Y., Zhou, S., Savarese, S., Xiong, C., Zhong, V., & Yu, T. (2024). OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments (No. arXiv:2404.07972). arXiv. https://doi.org/10.48550/arXiv.2404.07972

• We trained a multinomial logistic regression model (k = 18, n = 1.5 million) using racial probability classes extracted from DeepFace’s racial classifier and first and last name racial summary statistics from the US census and Rosenman et al. (2023), achieving 92.76% accuracy in three-race classification (Black, White, Hispanic).

• A sufficiently accurate linear model trained on biased data learns the true signal from noise. Systematic deviations indicate mislabeling by authorities rather than model error.

• 29% of individuals predicted to be Hispanic were officially classified as White by Department of Corrections authorities.

• This pattern persisted at high model confidence (95-100%), where 22.4% of predicted Hispanics were still assigned as White.

• Correcting for misclassification increases Hispanic criminal record rates by 31%, decreases White rates by 6%, and decreases Black rates by 1%.

• Bias between other racial pairings was minimal and symmetrical (equal numbers of Blacks misclassified as White and vice versa).

• State-level analysis showed no correlation with political ideology (r = 0.21, 95% CI: -0.36 to 0.67, p = 0.472), indicating random administrative error rather than deliberate bias.

• The proportion of predicted Hispanics assigned White (r = -0.80, 95% CI: -0.95 to -0.38, p = 0.003, n=11) and the proportion of predicted Whites assigned Hispanic both correlated with Native American ancestry among Latinos (r = 0.74, 95% CI: 0.26 to 0.93, p = 0.009, n=11).

Introduction

Every so often on X a viral post emerges showcasing how non-white races are being systemically misclassified as White.

The implication is that it never occurs the other way around and therefore must be malicious, a phenomenon of anti-white racism to cook the books in favor of non-whites, an attempt at reverse racism to hide that pesky 13/52.

It’s popular enough now that various anons have spawned numerous similarly viral collages. Here’s a small, roughly sewn compilation:

However, these accounts have issues that trigger skepticism among those who wouldn’t like finding out that non-white groups are more criminal than catalogued, and ambiguity for those inclined to accept the premise:

1. You have to trust the compiler doesn’t have ideological (right-wing) bias subtly pushing them towards cherry-picked cases.

2. Even ignoring 1, manual inspection has inherent issues. There’s variation in how someone classifies someone as White or not.

3. It’s anecdotal, limited to manual inspection. You have to trust that the n cases reviewed are representative, which is almost impossible to prove with this method.

There are likely more flaws than this. Everyone knows this isn’t evidence of systematic misclassification. Which is exactly the problem.

Luckily, we conducted extensive, systematic web-scraping of Department of Corrections (DOC) databases nearly a year ago, obtaining records for 5.5 million criminals from 39 U.S. states. After applying data quality requirements, our final dataset comprised 1.5 million criminal records with complete mugshot, demographic, and naming information.

We have the materials to answer this question definitively. Using statistical and modeling methods incorporating 21 variables derived from mugshots using facial recognition algorithms and names using demographic name databases (US census, Rosenman et al., 2023), we also have the means.

Is Bias Common?

Before we tackle the dataset and explain what we did, we must ask: is bias common in the criminal justice system?

This is important because if it is common and anti-white, finding bias in criminal racial classification would be unsurprising, or even suspicious if we didn’t find it.

On the other hand, if the criminal justice system is mostly color blind or genuinely pro-white, finding misclassification against Whites would be surprising.

So what does the literature say?

Racial bias has been studied across every stage of America’s criminal justice system, from initial police contact to final sentencing (Last, 2020). Experimental and observational studies spanning multiple domains with rigorous controls find either no racial bias against minorities or bias favoring minorities over Whites. Some prominent examples:

• The first randomized controlled experiment on prosecutorial bias tested nationwide prosecutors with realistic case vignettes and found no statistically significant relationship between defendant race and charging decisions, with some analyses showing pro-black treatment (Robertson et al., 2019).

• Experimental research using realistic deadly force simulators found officers were slower to shoot armed Black suspects than armed White suspects, and less likely to shoot unarmed Black suspects despite showing implicit bias (James et al., 2016).

• The largest jury study ever conducted analyzed 300,000+ felony cases over 32 years and found no taste-based or statistical discrimination against Black defendants, with similar disparate impact when race was unknown to jurors (Hoekstra et al., 2023).

• Comprehensive analysis of sentencing data across 43 U.S. states found either no racial bias in sentence length or bias favoring minorities, with studies consistently showing that legal factors (offense severity, criminal history) rather than race determine sentencing outcomes (Averages, 2023).

• Race becomes unrelated to arrest probability after controlling for IQ, impulsivity, and criminal history in samples of 1,331 ex-convicts and other populations (Schwartz et al., 2019; Beaver et al., 2013).

• Direct measurement using cameras found that the proportion of speeding drivers who were Black mirrored the proportion of Black drivers stopped by police (Lange et al., 2005).

• Using the rate of attacks on police as a benchmark, Black Americans were 40% less likely to be shot by police than White Americans (Shjarback & Nix, 2019).

• Multiple studies comparing arrest rates to incident reports found either no racial bias or pro-black bias in arrests, with consistent findings across 22 crime types (D’Alessio & Stolzenberg, 2003; Beck, 2021; Rubenstein, 2016).

• Analysis using violent crime rates as a benchmark found White people over-represented among police killings, with no evidence of anti-black bias in most estimates (Cesario et al., 2018).

So generally there’s either no bias, or in occasional cases mild to moderate pro-black bias. We can establish that criminal misclassification bias against Whites is a possibility, despite explicit anti-white bias not being systematic in the literature.

Data

Quality and Quantity

Every US state maintains a Department of Corrections, and most operate a functioning website allowing public access to criminal records.

Using web scraping technologies, we collected complete databases from all 51 jurisdictions.

Our bias analysis required specific data elements from each state. We established minimum requirements for inclusion:

1. Mugshots

2. Racial classifications, with White, Hispanic and Black as distinct categories

3. Complete names (first, middle, last) and suffixes

Point two is surprisingly restrictive. We excluded 16 states for lacking either race or mugshot data, and an additional 9 states for not recording Hispanic as a distinct racial category. Hispanic in the US isn’t considered a race but an ethnicity. As a result, some states opted out of classifying Hispanics entirely, instead classifying almost all Hispanics as White.

Since we’re interested in bias where race is consistently recorded, we removed states that didn’t explicitly and exclusively classify Hispanic as a category. A summary of the states scraped and exclusion criteria can be seen below.

Furthermore, data quality varied dramatically across states in terms of quantity, completeness, and accessibility.

North Carolina exemplifies comprehensive data availability, allowing full database downloads with detailed records dating to the 1970s, yielding 1.2 million unique criminal records. In contrast, California, despite having four times North Carolina’s population and similar incarceration rates, produced nearly ten times fewer accessible records after scraping. The figure below shows the final representation of criminal records by state.

Race

We identified seven consistent racial classifications across our state datasets: Black, White, Hispanic, Asian or Pacific Islander, Native American, Multiracial, and “Unknown or Other”. While states used varying nomenclature (e.g., “hispanic or latin american”, “mexican american”, “mexican national” for Hispanic categories), we standardized these classifications for analysis. All qualifying states recorded Black, White, Hispanic, Asian or Pacific Islander, and Native American categories, with the sole exception of Florida, which did not assign Native American classifications.

Mugshots

We processed mugshots using DeepFace (Serengil & Ozpinar, 2024), a comprehensive facial recognition and classification framework. The system processes each image through facial detection, cropping, and alignment before applying neural network classifiers.

DeepFace’s internal model generates predictions across six racial categories: Asian, Indian, Black, White, Middle Eastern, and Latino Hispanic. These racial categories do not align perfectly with the classifications used by Departments of Corrections. We could not use this as an independent model for verifying our results. Instead, we incorporate these predictions as variables in our custom classification model rather than using DeepFace’s outputs directly.

Sex

Preliminary analysis showed that including sex as a predictor variable provided no significant improvement to our model. We excluded sex from subsequent models to maintain parsimony.

Names

We extracted complete name information (first, middle, last names, and suffixes) for each individual from Department of Corrections records.

To predict race from names, we utilized the comprehensive dataset from “Race and ethnicity data for first, middle, and surnames” (Rosenman et al., 2023), which provides racial breakdowns for each name segment across five categories: White, Black, Hispanic, Asian Pacific Islander, and “other” race.

This primary dataset yielded 15 predictive variables (5 races × 3 name segments). We supplemented this with US 2010 census data (U.S. Census Bureau, 2010) for additional demographic breakdowns by race for surnames. The census data included two categories not present in the Rosenman dataset: “American Indian” and “2 or more races”, adding two more variables specifically for last names.

Since no comprehensive suffix dataset exists, we created boolean indicators for the most common suffixes: JR/SR designations and Roman numerals.

Preliminary analysis revealed that both middle name and suffix variables contributed insignificantly to classification accuracy. Middle names alone achieved 41.84% accuracy in five-race classification and 49.03% in three-race classification, worse than the naive baseline of predicting all individuals as White (54.46% and 55.7% respectively).

Suffix variables added to first and last names provided no significant improvement in either classification scheme. When both variable sets were removed from the full model (including mugshot and census features), the effect on accuracy was insignificant.

Additionally, principal component analysis showed middle name variables introduced spurious double clustering artifacts without improving racial separability. We excluded all middle name and suffix variables from subsequent analyses.

Overall, our final name-based prediction system comprised 12 variables: 5 from first names and 7 from last names. Combined with DeepFace’s 6 mugshot-derived variables, our complete model incorporated 18 predictive features for racial classification. The figure below highlights the pipeline for our model using these variables.

Method

Overview

Our goal is to identify systematic bias in racial assignment by corrections authorities. We do this by training a linear (not literally OLS, just not non-linear) model on DOC-assigned race labels, then interpreting systematic deviations between model predictions and official assignments as evidence of mislabeling rather than model error.

Why does this work?

A model trained on biased labels will still learn the underlying signal (true race) rather than the bias itself. This occurs when two conditions hold:

1. The predictors capture genuine racial differences (facial features, name demographics).

2. The model achieves sufficiently high accuracy that it fits the dominant signal in the data rather than noise.

Linear models are particularly well-suited for this task because they fit overall linear relationships between predictors and outcomes, effectively averaging through systematic bias patterns rather than overfitting to them. We opted to use Multinomial Logistic Regression. Under these conditions, systematic deviations between model predictions and official classifications indicate bias in the original labels. In simulations later, we validate this approach by introducing bias into a generated dataset and demonstrating that our method successfully detects it.

When training the model, it’s also critical that the prediction value of each class is weighted inverse to its proportion in the dataset.

This prevents our model from biasing towards classifying everyone as White or Black simply because Whites and Blacks constitute the majority of criminals. Without this weighting, it would appear that minority classes are mislabeled as White or Black, which is not what we’re trying to capture.

What Races Should we Include in our Analyses?

We don’t just have Black, White and Hispanic classifications. We also have Unknown, Other, Asian and Native American. This leaves us with many races we can predict, not just the three.

However, as discussed in the overview, we need sufficiently high accuracy to be confident that deviations capture mislabeling rather than model error. Thus, we should test how our model performs with everything (including Unknown + Other), the five-race classification, and the BWH (Black-White-Hispanic) combination only. We’ll determine for ourselves what is sufficient, but >90% accuracy on multi-class classification is preferable.

Separability and Dimensionality Reduction

A critical methodological concern involves disentangling genuine bias from natural classification difficulties arising from differential phenotypic distinctness between racial groups. Certain racial categories exhibit closer genetic ancestry, physical appearance, and naming patterns than others. Most notably, the relatively modest differences between Hispanics and Whites compared to the more pronounced distinctions separating these groups from Blacks.

This natural variation in inter-group distinctness generates a testable prediction: if misclassification patterns simply reflect inherent classification difficulty rather than systematic bias, we should observe higher accuracy rates for more phenotypically distinct groups (such as Blacks) and proportionally higher error rates for phenotypically similar groups (Hispanics and Whites). Under this scenario, elevated misclassification rates between similar groups could be attributed to reduced Euclidean distance in feature space rather than systematic bias.

This analysis is complemented by classification models, as they are fundamentally designed to maximize separability between groups, trained explicitly to find the boundaries that best distinguish categories even when those boundaries are subtle. If our model achieves high overall accuracy, it demonstrates that the racial groups are sufficiently separable in our feature space. Persistent misclassification of a specific pairing at high confidence would then suggest something other than mere phenotypic similarity—it would indicate cases where the model has detected strong, convergent evidence across multiple predictive features (mugshot analysis, name demographics) pointing toward one classification while official records show another.

When such high-confidence predictions contradict official classifications at rates exceeding random error, this indicates consistent patterns of misclassification rather than model uncertainty. We’ll explore the relationship between model confidence and racial classification to quantify whether these patterns reflect random errors or systematic bias.

Simulations

To validate our methodological assumptions and develop a framework for interpreting real-world bias patterns, we constructed controlled simulation studies using synthetic datasets with known bias characteristics. These simulations enable us to test whether our analytical approach can successfully detect and characterize different types of systematic bias under controlled conditions.

Our simulation framework employed a simplified three-group structure (Red, Blue, Green) designed to mirror the essential characteristics of our real-world three-race analysis while maintaining interpretive clarity. This design facilitates direct comparison between simulated and observed bias patterns.

We constructed our synthetic data within a two-dimensional feature space. Since our primary methodology relies on linear modeling techniques that do not assume complex nonlinear interactions between variables, higher-dimensional representations would introduce unnecessary complexity without enhancing our ability to detect the linear bias patterns central to our analysis. The two-dimensional approach also enables clear visualization of bias patterns and model performance.

Within this simplified space, we positioned each group with equal separation distances of three standard deviations between all group centroids, ensuring balanced distinctness across categories. We then introduced three distinct bias types affecting only the Blue and Green groups, leaving Red as an unbiased control:

• Random bias: Greens randomly assigned to Blue.

• Strategic bias: Greens closest to the Blue mean assigned to Blue.

• Obvious bias: Greens furthest from the Blue mean assigned to Blue.

For each bias scenario, we reassigned 10% of Greens to Blue classification, with the selection mechanism varying according to the specific bias type being simulated.

If you’re interested in the mathematics and setup, see the appendix.

State Analysis

Operating under our methodological assumption that model predictions represent ground truth, we calculated state-specific misclassification rates for each racial group, focusing on categories where systematic bias appears likely based on our overall findings.

To test whether any detected bias reflects deliberate discrimination versus random administrative error, we examined correlations between state-level misclassification rates and political ideology, measured through Republican vote share in recent elections. Systematic correlations with political variables would suggest intentional bias, while their absence would support alternative explanations such as administrative inconsistency or measurement error.

We also investigated whether genetic ancestry composition influences misclassification patterns. Using ancestry data from Bryc et al. (2015), we tested whether misclassification rates correlate with population-level genetic distinctness between racial groups within states. This analysis helps distinguish between bias arising from genuine classification difficulty due to genetic similarity versus systematic administrative bias.

Results

The Races We Included in our Analyses

Finally, the results! First, how do our models perform for each racial category?

The correlation matrix shows that the best model for five-race classification struggles to predict Native American and Asian or Pacific Islander. Native American achieves only 14% accuracy, Asians only 51%.

On the PC1/PC2 map, Native Americans and Asian Pacific Islanders confirm this challenge. The Native Americans cluster almost on top of Whites and overlap significantly with Hispanics.

Curiously, you can also see two distinct clusters for Asians and Asian Pacific Islanders. Isolating this cluster, we can examine the names of these individuals. Here’s a randomly selected sample of Asian Pacific Islanders within the top right cluster and outside of it. Indians and East Asians appear to belong to it.

Continuing on, Unknown + Other fared even worse. The model tended to ignore the classification, with predictions falling randomly to each race. We’ll dismiss these classes from now on.

A table looking at overall accuracy by the races we chose to predict shows that Black-White-Hispanic is the only model that achieved overall accuracy >90%.

This >90% threshold holds true for all races in the BWH model only. As a result, our analyses will focus on Blacks, Whites and Hispanics going forward.

All features contributed cumulatively and significantly to our result: both names and mugshots are important for accurate classification.

Flipping the Matrix

The correlation matrix above assumes that the assigned race is the true race and the predicted race is error. However, given the model assumptions we outlined, this is reversed! The assigned race is the potential error, with the true race being the prediction. We must flip the matrix:

Now we can see the misclassifications! Nearly 29% of genuine Hispanics were classified as White! This is a preliminary indication of bias, but it could also be the model failing. Phenotypically, Hispanics do look pretty similar to Whites after all. In the next section we debunk this.

Separability and Dimensionality Reduction

Here, we will show (1) in the majority of these cases the model is supremely confident that the Hispanic is “White” and (2) from dimensionality reduction, Blacks, Hispanics and Whites are equidistant, establishing that the races are separable and not closely related statistically as their appearance alone (Hispanics looking similar to Whites) would imply.

Starting with dimensionality reduction. This figure draws from a random sample of 10,000 Whites, Blacks and Hispanics from our dataset.

At a glance, these three groups appear about equally separable from each other. Computing the Euclidean distance across the first 16 PCs, Whites are actually closer to Blacks than to Hispanics, but Hispanics are closer to Whites than to Blacks, which confirms what the plot shows.

Zooming in on the PC plot, one can see a spattering of Whites in the Hispanic zones, but not necessarily vice versa. One could say these look like Hispanics mislabeled as White!

We can take a closer look at this. To do this, we plot just the Hispanics, then a contour of the White region. Then we plot the same directly below but in reverse to see the contrast. To make the plot clearer, we’ve added 50k Whites and Hispanic points (instead of just 10k).

Now that’s much clearer. There are some Hispanics in the White areas, but the reverse is far more prevalent—there’s a long tail of “Whites” in the Hispanic areas, far from the White zones. Crucially, the mix of Whites within the Hispanic distribution parallels the Hispanic distribution, whereas this pattern does not hold for Hispanics located in the White region.

But this is theoretical separability. How does a model, when appropriately trained on all our variables, distinguish these racial groups? Is the model also finding “Whites” far from the White zones? Yes, and many of them.

Looking at the following figure, we can see that as the model increases in confidence for its predictions of Blacks and Whites, it converges on a 100% match between the assigned race and the predicted race.

To read this plot: the title of each facet is the predicted race, the color aesthetic denotes the proportion by the assigned racial class.

As one can see, even when the model is 95-100% confident that a predicted Hispanic is Hispanic, 22.4% are labeled as White!

A skeptic might point out that high confidence cases might only constitute a tiny fraction of the overall misclassification. This is incorrect. The opposite is true. The median model confidence is 91.7% for predicted Hispanics assigned White.

Table X: Misclassification Confidence Statistics

Visualizing the distribution also provides context. First, predicted Hispanics have the same bias distribution not just for Whites but also for Blacks, although the latter case is orders of magnitude smaller. Second, for every distribution there’s a spike in cases where model confidence is >90%, indicating that there are severe misclassification examples for almost every race.

So far, these results alone would indicate, statistically, that there is indeed systematic misclassification of Hispanics as White.

Simulation

So what type of bias does this most closely reflect?

Recalling from our simulation method, we developed and tested three types of bias using the same method we just employed:

• Random bias: Greens randomly assigned to Blue.

• Strategic bias: Greens closest to the Blue mean assigned to Blue.

• Obvious bias: Greens furthest from the Blue mean assigned to Blue.

Here’s what they look like:

Looking at the results, our real-world results most closely reflect random bias. It has a high resting level of Blues (Whites) misclassified as Greens (Hispanics). It does not resemble obvious bias whatsoever. That is to say, comparing this to our real world data, there isn’t evidence that very obvious Hispanics are being systematically and exclusively assigned as White.

State Level Analyses

Now that we’ve documented bias from the perspective of separability and simulations, where a significant proportion (29%) of predicted Hispanics are being assigned White, we ask what predicts misclassification severity.

First, we look at ideology and variation in misclassification by state.

Misclassification of Hispanics as White varies greatly by state. As many as >60% of Hispanics in Florida are being assigned as “White”! This seems incredible.

Digging through the literature, Latino racial self-identification varies substantially by national origin, with research showing approximately 91% of Cuban Americans self-identify as White compared to 56% of Puerto Ricans and 49% of Mexicans (Michael & Timberlake, 2007; Figuereo et al., 2025). Given that Florida has a large Cuban population, it’s possible that this reflects self-identification as White manifesting in official records.

This might explain Florida being an outlier, but what about the other states? We can test for ideology using Republican vote share as a proxy. There is no significant association (r = 0.21, 95% CI: -0.36 to 0.67, p = 0.472). If anything, the above cited research suggested that Latinos were more likely to identify as White when living in Republican areas, although this was only tested on a regional basis. The expected direction here is positive, if there is one at all.

What about genetic similarity between Hispanics and Europeans? Hispanics in states with greater European heritage and less Native American ancestry might be confused with Whites more often. Crucially, we’re not interested in measuring European ancestry among Hispanics, but the distinctive Native American ancestry.

This would have direct causal effects. One might consider that the less distinctive Hispanics become from Europeans (less Native American ancestry), the harder it would be to distinguish them, the more the lines would blur between the ethnicities, commensurately increasing misclassification. We sourced the admixture data by state from Bryc et al. (2015).

Despite the small sample, it’s statistically significant in two directions: predicted Hispanics assigned White being negatively correlated (r = -0.80, 95% CI: -0.95 to -0.38, p = 0.003, n=11) and predicted Whites assigned Hispanic positively correlated (r = 0.74, 95% CI: 0.26 to 0.93, p = 0.009, n=11).

One could interpret these multiple ways:

1. The Boring Interpretation:

The hypothesis is correct. Native American ancestry causes Hispanics to become more distinctive and less likely to be misclassified as White. We can ignore the White-Hispanic correlation because it’s highly leveraged by Arizona and Colorado and the p-value is borderline anyway. This is reinforced by the fact that the ancestry data is based on low-n 23andMe samples, which are unfortunately selected, as it’s plausible that higher SES groups are more likely to use ancestry services, among other reasons.

2. The Less Boring Interpretation:

It’s about self-identification, not phenotypic distinctiveness through genetics. As we showed with Cubans in Florida, misclassification occurs when Hispanics relative to other states choose to identify as White despite their surnames, appearance, etc.

Given this premise, we know our model already incorporates phenotypic distinctiveness. It extracts ethnicity from DeepFace, and darker skin among Latinos robustly predicts identifying as such (Michael & Timberlake, 2007; Figuereo et al., 2025). So our model should already control for phenotype, excluding the possibility of inter-state variation in misclassification resulting from appearance deriving from admixture.

Yet Native American ancestry among Latinos does, so the alternate hypothesis is that Hispanic identification is more likely controlling for phenotype and surname where it’s more visible. There could be group effects to Hispanic identity: when it becomes more distinctive and visible, others are more likely to recognize it and identify the same. This would explain why Whites are over-identifying as Hispanic (the positive correlation) and why misclassification of Hispanics as White is rarer (the negative correlation).

So which interpretation is correct?

To test this further, we added state to the MLR classification model as a control for variation in self-identification. If the self-identification hypothesis were correct, adding state should meaningfully improve model performance. State would act as a proxy for regional self-identification norms (Cubans in Florida identifying as White, different patterns in states with varying Native American ancestry), allowing the model to learn state-specific patterns like “in Florida, people with X features are more likely to identify as White.” This should help the model match official classifications better.

However, it did not help. The classification accuracy only increased by ~1% and the mass of Hispanic misclassification remained. This suggests that state-level self-identification patterns aren’t actually explaining the misclassification—if they were, state should have been a powerful predictor.

The data quality is simply not here to say for certain. It should be noted that other combinations of ancestry regressions were insignificant. For example, correlating European ancestry among Latinos to misclassification instead of Native American ancestry. So it’s also possible that this is just noise. Personally I’m biased towards Occam’s razor, the simplest explanation is often the best.

Collages of Mugshots: Individual Inspection

Despite the reservations we have about manual inspection, in the context of our model it’s still worthwhile as it validates our approach visually.

We used Python to randomly and programmatically select and render mugshot collages to verify our approach. First, we inspected low confidence and high confidence predictions. Contrasting these, we found that model confidence was commensurate with the predicted race being the likelier true racial classification, confirming our methodology.

Continuing on, we then generated mugshots for the strongest and weakest misclassification directions: predicted Hispanics assigned “White” and predicted Blacks assigned “White”. The median confidence for predicted Hispanics assigned White was 91%, whereas for predicted Blacks assigned White it’s only 51%.

Observing the two, it is apparent from the surnames and physical appearances that predicted Hispanics are indeed Hispanic instead of White. In contrast, for predicted Blacks assigned White it is clear that the discrepancy between predicted and assigned race results from model error. Once again, this confirms that our methodology is sufficient and that Hispanics assigned White is systematic.

Correcting the Bias: What are the True Crime Rates?

To quantify the impact of racial misclassification on reported criminal record rates, we corrected the assigned racial counts in each state’s criminal database. We calculated DOC criminal records per 100,000 population using census data, normalizing all rates relative to the assigned White rate within each state.

We applied two corrections:

First, a high-confidence correction. We assumed that the predicted race was the true race where the model classified with >90% confidence; for cases where model confidence was <90%, the assumed race was the assigned race. Second, a more generous reclassification. We assumed that the predicted race reflected the true race rather than the assigned race for all cases. One can see the results by state and race.

Overall, the second adjustment correcting for misclassification increases Hispanic criminal record rates by 31% among the states analyzed while decreasing Black rates by 1% and White rates by 6%.

The high-confidence adjustment was more modest. Hispanic criminal record rates still increased by 20%, however Black criminal rates only fell an indistinguishable 0.2%, and White rates fell by 4%. The figure below shows the combined criminal record rate adjustment by state and overall.

How does this affect the overall rates? Since we’re working with criminal records of individuals that vary by state, not incarceration rates, we normalized the assigned White rate to 1 as a reference point. Post-adjustment, White and Hispanic criminal record rates converge.

These are not to be interpreted as crime rates—these are changes in the rate of records for all states combined. These records stretch back decades when the population composition was different, which would affect the denominator in this calculation. This is the reality of the data we’re working with. This is why the adjustments were visualized first to avoid confusion.

These results demonstrate that racial misclassification in criminal databases artificially deflates Hispanic criminal record rates while mildly inflating White rates. The magnitude of the effect is substantial: the corrected Hispanic rate is 20% higher even in the conservative case.

Conclusion

Four findings support genuine Hispanic-to-White misclassification rather than model error:

1. Visual inspection of high-confidence cases showed individuals with Hispanic phenotypes and surnames despite White classifications.

2. Principal component analysis revealed a separability paradox: Whites and Hispanics showed greater Euclidean distance (4.53) than Whites and Blacks (3.41) in PC space, yet Hispanics were misclassified as White at high rates while Blacks were not. While PCA separability differs from predictive separability, this pattern contradicts explanations based solely on phenotypic similarity. Additionally, Whites were distributed throughout the Hispanic region in PC space, but Hispanics were not distributed throughout the White region, suggesting that individuals classified as White within Hispanic PC space represent misclassified Hispanics.

3. Model confidence predicted accuracy for Black and White classifications: as confidence approached 100%, accuracy approached 100%. For Hispanic predictions, this relationship broke down. At 95-100% confidence, 22.4% of predicted Hispanics were still assigned as White. This pattern was asymmetric, occurring only for Hispanic-to-White misclassification at high rates (>20% in most states), not for other racial combinations. When we simulated datasets with known random label bias (randomly reassigning 10% of one group to another), models trained on the biased labels reproduced this exact pattern: normal confidence-accuracy relationships for unbiased groups, but a high error ceiling equal to the bias prevalence for the biased group. The real-world data showed the same signature, indicating label bias rather than model limitations. This pattern was not an artifact of class imbalance (more Whites than Hispanics in the dataset) because we used inverse frequency weighting to ensure the model optimized for balanced performance across all racial groups.

4. Simulations aligned with what we found: a random distribution of assignments of Whites (Blues) in the Hispanic space (Greens).

The viral posts were right that Hispanics are being misclassified as White in criminal records—29% of them, in fact. But the misclassification appears random rather than deliberate. State-level patterns show no correlation with political ideology, and the distribution matches what we’d expect from administrative error rather than systematic bias. Correcting for this increases Hispanic criminal record rates by 20-31% and decreases White rates by 6%, narrowing the gap between the two considerably. The anecdotal collages got people asking the right question. We just needed 1.5 million records and a multinomial logistic regression to answer it.

Replicate: Open Data

OSF

OSF all the supplementary materials here:

• Copy of the article in paper format. It has more academic language and is slightly more comprehensive.

• The GitHub repository as a .zip

• The full dataset as a .zip excluding mugshots (as these are hundreds of GBs).

GitHub

The GitHub repository containing the code generating this full analysis is open source. To run this repository with the data below, you’ll need to access the data (on OSF), download it, unzip it in the same folder/directory, start the Python environment, install dependencies if you haven’t (R and its packages), then run ./replicate.sh. There’s a README.md you can follow.

Akiba, T., Sano, S., Yanase, T., Ohta, T., & Koyama, M. (2019). Optuna: A next-generation hyperparameter optimization framework. \*Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining\*. https://optuna.org/

Averages. (2023, November 6). Do minorities get longer sentences? An analysis of every state. \*The Law of Averages Blog\*. https://thelawofaveragesblog.wordpress.com/2023/11/06/do-minorities-get-longer-sentences-an-analysis-of-every-state/

Beaver, K. M., DeLisi, M., Wright, J. P., Boutwell, B. B., Barnes, J. C., & Vaughn, M. G. (2013). No evidence of racial discrimination in criminal justice processing: Results from the National Longitudinal Study of Adolescent Health. \*Personality and Individual Differences\*, \*55\*(1), 29-34. https://doi.org/10.1016/j.paid.2013.01.020

Beck, A. J. (2021). \*Race and ethnicity of violent crime offenders and arrestees, 2018\* (NCJ 255969). Bureau of Justice Statistics. https://bjs.ojp.gov/content/pub/pdf/revcoa18.pdf

Bryc, K., Durand, E. Y., Macpherson, J. M., Reich, D., & Mountain, J. L. (2015). The genetic ancestry of African Americans, Latinos, and European Americans across the United States. \*American Journal of Human Genetics\*, \*96\*(1), 37-53. https://doi.org/10.1016/j.ajhg.2014.11.010

Cesario, J., Johnson, D. J., & Terrill, W. (2019). Is there evidence of racial disparity in police use of deadly force? Analyses of officer-involved fatal shootings in 2015–2016. \*Social Psychological and Personality Science\*, \*10\*(5), 586-595. https://doi.org/10.1177/1948550618775108

D’Alessio, S. J., & Stolzenberg, L. (2003). Race and the probability of arrest. \*Social Forces\*, \*81\*(4), 1381-1397. https://doi.org/10.1353/sof.2003.0051

Figuereo, V., Rosales, R., Takeuchi, D. T., & Calvo, R. (2025). What race am I? Factors associated with racial self-classification among U.S. Latinx adults. \*Du Bois Review: Social Science Research on Race\*, \*22\*(2), 215-236. https://doi.org/10.1017/S1742058X25000050

Hoekstra, M., Oh, S., & Tangvatcharapong, M. (2023). \*Are American juries racially discriminatory? Evidence from over a quarter million felony grand jury cases\*. Texas A\&M University Department of Economics. https://economics.ucr.edu/wp-content/uploads/2023/10/10-30-23-Hoekstra.pdf

James, L., James, S. M., & Vila, B. J. (2016). The reverse racism effect: Are cops more hesitant to shoot black than white suspects? \*Criminology & Public Policy\*, \*15\*(2), 457-479. https://doi.org/10.1111/1745-9133.12187

Lange, J. E., Johnson, M. B., & Voas, R. B. (2005). Testing the racial profiling hypothesis for seemingly disparate traffic stops on the New Jersey Turnpike. \*Justice Quarterly\*, \*22\*(2), 193-223. https://doi.org/10.1080/07418820500088952

Last, S. (2020, June 3). American racism and the anti-white left. \*Ideas and Data Blog\*. https://ideasanddata.wordpress.com/2020/06/03/american-racism-and-the-anti-white-left/

Michael, J., & Timberlake, J. M. (2007). Are Latinos becoming white? Determinants of Latinos’ racial self-identification in the United States. \*Social Forces\*, \*86\*(2), 561-573.

Robertson, C. T., Baughman, S. B., & Wright, M. (2019). Race and class: A randomized experiment with prosecutors. \*Journal of Empirical Legal Studies\*, \*16\*(4), 807-847. https://doi.org/10.1111/jels.12235

Rosenman, E. T., Olivella, S., & Imai, K. (2023). Race and ethnicity data for first, middle, and surnames. \*Scientific Data\*, \*10\*(1), 1-16. https://doi.org/10.1038/s41597-023-02202-2

Rubenstein, E. S. (2016). \*The color of crime: 2016 revised edition\*. New Century Foundation. https://www.hoplofobia.info/wp-content/uploads/2021/02/2016-The-Color-of-Crime.pdf

Schwartz, J. A., & Beaver, K. M. (2019). A longitudinal examination of the association between intelligence and rearrest using a latent trait–state–occasion modeling approach in a sample of previously adjudicated youth. \*Developmental Psychology\*, \*55\*(10), 2121-2135. https://doi.org/10.1037/dev0000838

Serengil, S. I., & Ozpinar, A. (2024). A benchmark of facial recognition pipelines and co-usability performances of modules. \*Journal of Information Technologies\*, \*17\*(2), 95-107. https://doi.org/10.17671/gazibtd.1399077

Shjarback, J. A., & Nix, J. (2019). Considering violence against police by citizen race/ethnicity to contextualize representation in officer-involved shootings. \*Journal of Criminal Justice\*, \*65\*, 101653\. https://doi.org/10.1016/j.jcrimjus.2019.101653

U.S. Census Bureau. (2010). \*Genealogy data\*. https://www.census.gov/topics/population/genealogy/data/2010_surnames.html

For the Appendix, visit my site’s post, scroll to the bottom.

TL;DR

• A positive eugenics policy ($20,000 annual child benefit per standard deviation of parental IQ above population mean) produces 2-18 IQ point gains over a century, depending on policy effectiveness

• GDP per capita increases 22% to 6.5-fold across scenarios, with the high-effectiveness scenario reaching $242,559

• All scenarios experience catastrophic population decline given humanity's demographic crisis. With the no-policy control losing 70% of its population.

• Base fertility rates (what occurs if policy is abolished) collapse to 0.66-1.14 children per woman by year 100, creating permanent policy dependency

• Policy cost equals 3.2% of GDP annually if implemented in the US; equivalent to the US military budget.

• Innovation index starts from Poland/Greece/Ireland levels to Switzerland/USA/Sweden.

Introduction

In our previous article we calculated the effect of population decline and dysgenics on innovation.

Naturally, such inquiry also raised another interesting question by Joseph Bronski: if eugenics had been employed decades earlier, how much more innovative progress would we have had?

Would we have developed LLMs and the internet in the 1960s? Would the singularity have already occurred by now? When would we have landed on the moon?

However and alas, the effect of eugenics and the effectiveness of its policies is complex. There are many varying types of policies one could implement, each with their pros and cons.

For example, broadly speaking, negative eugenics would reduce the overall population of a country through reduction of fertility. Likewise, positive eugenics could fail, given that there is no known tested policy intervention that has successfully elevated the fertility of any population beyond rounding error, let alone a target demographic. Furthermore, any policy successful at increasing the IQ of a country would proportionally decrease its fertility, given its correlation with national IQ.

To help ameliorate the complexities of such hypothesized effects, we employ a series of simulations.

We generate our own novel population projections, modifying the fertility, IQ, innovation and other statistics of a million persons, proportional to the effectiveness of a eugenics program.

The effect sizes of such policies are determined by what we know about fertility altering pro-natal policies. We use this in addition to the employment of the UN's cohort component method and quantitative genetics on an individual level to generate a hypothetical positive eugenics policy outcome.

**Warning: the method ahead is for the nerdy. If you're more interested in the eugenics policy, its effectiveness, outcomes and other key assumptions, we recommend skipping to the section "Method: The Eugenics Policy".**

Method: Main Loop

From top to bottom, this is how our simulation works:

First, we generate the initial population of 1 million persons. For convenience, we use the exact same age and sex distribution as the United States as of 2023—this was provided by the UN.

Next, we need to assign IQs to these individuals. To simplify our simulation, we do not assume any sex differences in intelligence or differences by age. To accomplish this, we give everyone a genotypic IQ and an environmental IQ, then sum these to obtain their resultant phenotypic IQ.

Let \\(P\\) be the phenotype; it is the sum of a genetic component \\(G\\) and an environmental component \\(E\\).

The total phenotypic variance in the population \\(V_P\\) is partitioned into the variance of the genetic values \\(V_G\\) and the variance of the environmental effects \\(V_E\\), assuming no covariance between genetic and environmental factors.

Heritability \\(h^2\\) is defined as the proportion of the total phenotypic variance that is attributable to genetic variance. For our simulation, \\(h^2\\) is key and is assumed based on the literature:

Given the defined total phenotypic variance \\(V_P\\) and heritability \\(h^2\\), the genetic and environmental variances are determined as follows:

We assume, according to standard IQ scaling, that the standard deviation of IQ is 15. Variance equals the standard deviation squared, so in this case:

Genetic Value (G): Sampled from a normal distribution with a mean equal to the population's starting mean phenotype \\(\mu_P\\) and a variance of \\(V_G\\).

Environmental Value (E): Sampled from a normal distribution with a mean of 0 and a variance of \\(V_E\\).

This results in the following initial code:

Next, now that we have the initial population, we need to begin aging the population by one year. This involves the following steps:

1. Coupling our simulated persons for assortative mating.

2. Generating the births resulting from these couples for that year.

3. Estimating who will die and at what age. This includes child mortality for the newborns.

Assortative Mating

To simulate assortative mating, we generate a "desired partner phenotype" for each fertile female. This desired phenotype is correlated with the female's own phenotype, controlled by the mating correlation parameter, which is set to \\(r=0.4\\).

This process ensures that females with higher phenotypic values will, on average, desire partners with higher phenotypic values, and vice versa.

The correlation value is sourced from a meta-analysis of twin studies.

Let \\(P_f\\) be the phenotype of an individual female. We first standardize this phenotype relative to the mean \\(\mu_{P,f}\\) and standard deviation \\(\sigma_P\\) of the fertile female population.

Next, we generate the standardized desired partner phenotype, \\(P_{d, \text{std}}\\). This new variable is constructed to correlate with \\(P_{f, \text{std}}\\) according to our assortative mating parameter. This is achieved by combining the female's standardized phenotype with a random component \\(\epsilon\\), where \\(\epsilon\\) is drawn from a standard normal distribution \\(\mathcal{N}(0, 1)\\).

Finally, this standardized desired phenotype is converted back to the original phenotypic scale to yield the final desired partner phenotype, \\(P_d\\).

All this is computed in the code section starting from population_fertile_females.

This next part is important.

These are only the *desired* partners, not the actual male partners that exist in our population. We conduct matching under two key constraints in our simulation:

1. Age-Restricted Mating: A female of a specific age can only be matched with males who are a fixed number of years older (three years in this simulation). This mimics realistic age-pairing patterns. The algorithm iterates through each fertile female age group (15-49) to find eligible partners.

2. Monogamous Pairing: Each male can only be part of one couple per year. Once a male is matched, he is removed from the pool of available partners for that year's mating cycle.

The matching itself is a deterministic, rank-order process designed to satisfy the females' preferences as closely as possible given the available males. For each age-based group:

1. The number of couples to be formed is determined by the size of the smaller group—either the females of that age or the eligible older males. A random sample of this size is drawn from both groups. Usually there are fewer females than males; however, so all females are paired.

2. The sampled females are then sorted in ascending order based on their desired partner phenotype.

3. The sampled males are also sorted in ascending order based on their own actual phenotype.

4. Couples are formed by pairing the individuals at the same rank in each sorted list. In other words, the lists are bound together. The female desiring the lowest-phenotype partner is matched with the available male who has the lowest phenotype; the female desiring the second-lowest is matched with the male who has the second-lowest, and so on.

Following from our previous section, this is coded in R as:

To verify this method worked, we can visualize the outcome of our algorithm below. It appears to show a correlation of 0.4, and when we verify this in R, we indeed achieve this value.

Births

To simulate which couples produce births, we need the Age-Specific Fertility Rate (ASFR). This is typically represented as the number of births per thousand women for that specific age.

As we did with our population structure, we can rely on the UN to provide the template for ASFR; again we chose the United States in 2023.

In our simulations, we use this value divided by one thousand, so we have the number of births per woman, instead of the number of births per thousand women.

This is crucial because we then treat the ASFR as a probability: the probability that a woman in a given year produces a child.

Using this approach, for all the couples that we assigned above, for that given year, we sample the couples where the probability of being sampled equals the ASFR of the woman. The remaining couples, post-sample, will all generate newborns.

Offspring Genetics, Phenotype and Sex

The genetic and phenotypic IQs of newborns are determined by a combination of parental genetics, random chance from meiosis (genetic variation between siblings), and environmental factors.

The inheritance process for each newborn is modeled in three main steps: determining the genetic value, adding environmental influence, and finally calculating the resulting phenotype.

A child's genetic value, \\(G_c\\), is derived from the genetic values of their parents, \\(G_f\\) (female) and \\(G_m\\) (male). The process begins by calculating the mid-parent genetic value, \\(G_{\text{mid}}\\), which is the simple average of the parents' genetic values.

This mid-parent value represents the expected genetic value of the offspring. However, sexual reproduction involves random chance in which alleles are passed on, a process known as Mendelian segregation. We model this by adding a random segregation component, \\(M_{\text{seg}}\\), drawn from a normal distribution with a mean of 0.

The variance for this segregation is half the total genetic variance \\(V_G/2\\) because the child inherits a random half of each parent's genes, and the variance of the mean of these two halves is \\(V_G/2\\). This term captures the genetic lottery that makes siblings different from one another.

The final genetic value of the child is the sum of the mid-parent value and the segregation component:

Just as with the initial population, each newborn is assigned an environmental value, \\(E_c\\), which is drawn independently from a normal distribution with a mean of 0 and a variance of \\(V_E\\).

This component represents all the non-genetic factors that contribute to an individual's phenotype. The child's final phenotypic IQ, \\(P_c\\), is the sum of their genetic and environmental values.

The sex of each newborn is assigned randomly, governed by the sex ratio at birth (SRB). The SRB gives the number of male births per 100 female births. The probability of a newborn being male, \\(P(\text{male})\\), is calculated as:

Each newborn is assigned "Male" or "Female" based on this probability.

The births, offspring traits and the eugenics policy (which we will discuss later) are modelled in R as follows:

Aging, Death, and Child Mortality

After accounting for new births, the entire population must be aged forward one year, and the chances of survival for each individual must be calculated. This process determines who dies and is removed from the simulation.

The model for mortality relies on age- and sex-specific survival probabilities derived from UN data. As with the initial population structure and fertility patterns, we use data for the United States in 2023 to ensure a realistic demographic baseline.

The core of this process is the life table parameter \\(p_x\\), which is the probability that an individual of age \\(x\\) will survive to their next birthday at age \\(x+1\\). These probabilities are provided separately for males and females.

The simulation of mortality proceeds as follows:

1. Aging: Each person's age is incremented by one year.

2. Death: Using these survival probabilities \\(p_x\\), simulated individuals are removed from the population. This represents their death.

Crucially, this survival mechanism also applies to the cohort of newborns. A newborn enters the simulation at Age = 0. Before the simulation proceeds to the next year, they face the infant survival probability, \\(p_0\\). Only the newborns who survive this initial check are added to the population for the following year.

The mortality for the population, male and female by age, including newborns, is modeled in R as follows. This final code snippet represents the end of our loop—one full year of simulation.

Method: The Eugenics Policy

Our eugenics policy is simple:

The IQ of everyone in the population is known through mass testing. When a couple produces a child, the government takes the average of both parents' IQ scores.

That couple's average IQ score is then compared against the general population for that year. "For that year" is crucial, since a population's average IQ, especially under a eugenics policy, can be changing year over year.

For every standard deviation (15 IQ points) the couple's IQ is above the population mean for that year, a child benefit of $20,000 USD is administered to the couple annually—inflation adjusted—with the last payment being made on the offspring's 20th birthday.

This policy is a *positive eugenics* policy; that is to say, couples with an average IQ below the population mean simply receive no benefit, instead of having to pay.

The payment is an addition to all present child benefits that may or may not exist for this hypothetical country or scenario. Additionally, it occurs regardless of marital status, couple living arrangements, or offspring death.

This last clause is to keep our scenario simple and applicable to all couples. Realistically, one would provide incentives to maintain and promote marriages and avoid rampant single-motherhood. However, modeling such effects is completely beyond the scope of our simulation.

Statistically, it's important to note that variance in the general population is less than the variance between couple averages.

This is because part of the variation cancels itself out when averaging two random people. In our sample, for example, the standard deviation in couple mean phenotypic IQ was 12.574 (this adjusts for assortative mating), not 15 as it is for the general population. As another example, for our couples, 11.7% had an average IQ greater than 115, but for any one person it is simply 15.9%. This is somewhat offset by assortative mating, which mitigates this phenomenon.

Policy Effectiveness

To determine the effectiveness of this policy on fertility for the targeted individuals in our simulation, we draw on Lyman Stone's article on pro-natal policies.

Appendix: Academic Research on Pro-Natal Policies

I have identified 34 academic studies since 2000, with the vast majority published since 2005, assessing the effectiveness of specific pro-natal policies. Of those studies, 22 contain sufficiently detailed estimates of effects and have sufficiently quantifiable costs, so that they can be used to estimate plausible elasticities of fertility with respect to various pro-natal incentives. The studies are listed in an appendix linked here. However, the summary finding is simple: an increase in the present value of child benefits equal to 10% of a household’s income can be expected to produce between 0.5% and 4.1% higher birth rates. The figure below presents 26 different elasticity estimates from the 22 studies with enough data to produce such an estimate.

The real median household income for the USA in 2023 was $80,610. Using this figure and Lyman Stone's findings, we can calculate the effectiveness of our positive eugenics policy.

We need to convert these figures into an effect per standard deviation. The mathematics for this is as follows:

Let \\(B_{\text{SD}}\\) be the final policy effect multiplier per standard deviation. This value determines the multiplicative fertility boost applied for each standard deviation a couple's average phenotype is above the population mean.

It is derived from several economic and effect-size parameters. Let \\(C_{\text{policy}}\\) be the total financial cost of the policy incentive per child, \\(I_{\text{household}}\\) be the median household income, and \\(E_{\text{base}}\\) be the baseline policy effectiveness, representing the fertility increase in percentage points for a financial incentive equivalent to 10% of household income.

The total percentage effect, \\(E_{\\%}\\), calculates the total fertility increase our policy is expected to generate. It first determines the policy's cost relative to a 10% slice of household income, and then multiplies this ratio by the base effectiveness.

The total percentage effect is then converted into a multiplicative factor, \\(B_{\text{SD}}\\), which will be used in the exponential model for the policy's dose-response.

Combining these steps into a single formula provides the final calculation for the policy effect per standard deviation.

With this calculation, we can determine the quantitative impact on fertility for our couples.

Crucially, we exponentiate the effect of the policy on fertility. For example, if a couple is 3 standard deviations from the mean, their fertility is multiplied by \\({B_{\text{SD}}}^3\\) instead of \\( (B_{\text{SD}} - 1) \times 3\\). This is a theoretical choice; TFR, when correlated to most outcomes—GDP per capita, NIQ, HDI, etc.—often has a higher correlation when it is log-linear. In layman's terms, fertility is relative—that is to say, a country going from a TFR of 5 to 3 is not analogous statistically to another country going from a TFR of 2 to 0.

For our simulation, this policy effect multiplier manipulates couples' fertility by raising the probability they will produce offspring. In R, this is coded as follows:

IQ x TFR Feedback Loop

There's one more modification we make to the simulation. Realistically, as our simulated country's standard of living, GDP per capita, and years of education increase causally commensurate with IQ, we should expect fertility to fall.

To incorporate this into our simulation, we set the TFR of our simulation equal to the predicted TFR in 2023 for a country with an NIQ of 100. Then, every year, we proportionally adjust the TFR of our country according to its change in NIQ. This is coded into our simulation as follows:

Policy Cost

For calculating the total cost of this policy, once again we decided to keep it relative to the USA since we used its household income to calculate policy effectiveness.

To estimate the policy cost, we created a shortened simulation up until the birth of the newborns. Then, we calculate the sum total of standard deviations above the average that the couples' phenotype was for those newborns.

The total sum of the standard deviations above the mean is then multiplied by the total expected cost the government would have to pay over the course of the entire policy for each standard deviation—that is, $400,000 USD (20 years × $20,000).

Since our population is scaled to 1 million, we multiply the cost by the US population in millions to convert it back to the US (343.4773 million as of 2023).

Results

Finally, the moment you've been waiting for: the results of our simulation. Given the large variation in the theoretical effectiveness of such policies, we ran the simulation four times. Once as a control with no policy effectiveness, and three times varying Lyman Stone's \\(E_{\text{base}}\\)—that is, the expected percent increase in births given a 10% increase in the present value of child benefits.

Remember that \\(E_{\text{base}}\\) was between 0.5% and 4.1%. The confidence intervals became the "Low" and "High" effectiveness scenarios respectively, with the average being the "Medium". Our simulation ran for 100 years for each scenario.

Policy Cost

The total estimated policy cost, if implemented in the US, totals approximately $886 billion for the first year. Taking the US GDP of 27.72 trillion in 2023, this is a cost of ~3.2% of GDP.

That is roughly what the US spends on the military and nearly half of what the US spends on education.

NIQ

As expected from the control scenario, IQ was a stable 100 beginning to end. By the end of the medium scenario, IQ had increased by an average of nearly 10 points.

Comparing this increase to other countries globally, this is equivalent to moving from a national IQ near the Yookay, the Netherlands, Finland, or China, to 2 IQ points higher than Singapore, the highest IQ country.

So how effective was the policy at boosting fertility by an individual's IQ? Well, this varies by the assortative mating in our population, given that our policy functions on couples, not individuals.

We quickly built a mini-simulation on our policy to calculate the effect by IQ and assortative mating values.

In our scenario, we assumed a correlation of 0.4, which is the second line from the bottom in this plot. So the average female in our simulation, with an IQ between 110-115, should have a TFR near replacement. Those couples with an IQ between 140 and 145 should expect to have at least 4 children.

It should also be noted that females with an IQ below average can still benefit from this policy! That is, if they happen to couple with a male with sufficiently high IQ to bring their average above the mean.

Demographics

Next, we have the effect of a pro-natalist policy on the demographics of our hypothetical million-person country. It should be noted that the TFR adjustment took into account the policy and the change in TFR as a result of the effect of changing NIQ over time.

By the end of the century, the TFR of our country had dropped below the baseline!

However, the positive effects on the overall population remained. The no-policy scenario resulted in a catastrophic loss of population. With its TFR near 1.2, nearly half the replacement rate, this is hardly surprising.

Economic and Innovative Effects

The resulting economic effects of our policy are miraculous for GDP per capita. At the start of our simulation, the GDP per capita was ~$37,600 (constant 2015 USD), equivalent to France, Japan, or Italy.

However, by the end of the medium scenario, it had reached ~$108,000, equivalent to Luxembourg, Bermuda, or ~160% of the USA. This growth is approximately an additional 1.06% compounding annual growth rate.

As for the total GDP, only in the radical "High" scenario did it result in total economic growth sufficient to offset its own population decline.

Finally, we applied our innovation index (this is a per capita metric) from our previous post on dysgenics. The initial 100 IQ population starts equivalent in innovation index to approximately Poland, Greece, or Ireland at 2.3 and ends with an innovation index similar to Switzerland, USA, or Sweden at 3.7. For reference, the most innovative country is Germany at an innovation index of 4.04.

Conclusion

Now, naturally, the question arises: would eugenics work? Given the large variation in theoretical effectiveness, our simulation provides a definitive answer—though perhaps not the one eugenic proponents might expect.

Demographics Summary

First, the cognitive effects are undeniable. Even the modest "Low" scenario produces a 2-point IQ gain over a century. The "Medium" scenario generates a substantial 10-point increase, while the "High" effectiveness scenario yields an 18-point improvement that would place a nation 2 IQ points above even Singapore, currently the world's highest-IQ country.

However, and this is important, the demographic story tells a more complex tale. All scenarios, including the control, experience catastrophic population decline. The no-policy scenario sees a 70% population collapse, while even our successful interventions witness 42-66% declines. Notably, the Base TFR—that is, the fertility rate that would prevail if the policy were suddenly abolished—reveals an alarming trend. By year 100, the Medium scenario's base TFR drops to 0.86, and the High scenario plummets to just 0.66. This means that the policy becomes not merely beneficial but essential for demographic survival; without it, fertility would crash to levels incompatible with civilization itself.

Economic and Innovation Summary

Next, the economic transformation is nothing short of miraculous. GDP per capita increases by 22% in the Low scenario, nearly triples in the Medium scenario, and increases by over 6.5-fold in the High scenario—transforming a population from middle-income status (comparable to France or Japan) to among the world's wealthiest nations. The innovation index similarly improves dramatically, suggesting accelerated technological progress that could fundamentally alter the trajectory of human development.

Crucially though, total economic output tells a different story. Despite the massive per-capita gains, total GDP actually falls in all scenarios except the most optimistic "High" scenario. While individuals become dramatically wealthier and more innovative, the shrinking population undermines total economic power until only the most effective policy implementation maintains both cognitive enhancement and demographic viability.

So, to return to our original question—would eugenics work? The answer is yes. The policy succeeds in enhancing intelligence and individual prosperity, yet it simultaneously creates dependency. At an estimated cost of 3.2% of GDP annually, the intervention represents a substantial but potentially justified investment, particularly given that the alternative—as shown by the collapsing base TFR—may be demographic extinction.

The most optimistic scenario suggests truly transformative benefits: a nation could become a technological and economic powerhouse within a century. However, this requires unprecedented policy effectiveness sustained over multiple generations, and the fundamental tension between cognitive enhancement and fertility presents challenges that extend far beyond selective breeding policies alone. Eugenics would work, but it would also become essential—a policy that, once implemented successfully, could never safely be abandoned. Of course a world with eugenics would yield a completely alternative political landscape. One such that may permanently find a solution to the demographics crisis we face.

TL;DR

• Frequency > Length: Posting more often (even daily) seems better than writing fewer, longer posts. Consistency in *timing* isn't key, but *volume* is.

• Price Matters: Higher average subscription prices strongly correlate with higher estimated revenue.

• Momentum is Real: A post's likes are overwhelmingly predicted by the average likes of the previous 10 posts (explaining ~86% of the variance!).

• Substack Boosts the First Post: Your first post gets a *huge* boost – make it count.

• Paid Post Sweet Spot: Aiming for roughly 50% paid posts appears optimal for maximizing revenue, though the relationship isn't perfectly linear.

• Category Counts: Culture, US Politics, and Finance Substacks tend to have higher revenue potential, while Fiction, Philosophy, and Travel lag behind in our model. For individual posts, Comics and Health Politics see the biggest like boosts relative to the baseline (Arts).

Introduction

Substack has exploded, becoming a go-to platform for writers, journalists, and creators looking to build direct relationships with their audience and monetize their work. But with thousands of newsletters vying for attention, what actually separates the breakout hits from the ones that fizzle out?

Is it brilliant prose? Niche topics? A relentless posting schedule? Or just plain luck?

As usual, rather than relying on anecdotes, we decided to dive into the data. We scraped information from a vast number of Substack publications and their posts, connecting it with backend data on pricing, subscriber counts (where available), and post statistics. We then built a couple of models to try and decode the patterns behind Substack success:

1. Predicting Substack Revenue: What publication-level factors (price, age, frequency, category, etc.) correlate with higher estimated earnings?

2. Predicting Post Likes: What makes an individual post resonate more with readers (length, paid status, timing, category)?

Let's see what the numbers tell us.

Method

To tackle this, we gathered data on posts (like counts, word counts, publish dates, paid status) and publications (subscriber estimates, pricing plans, categories, creation dates).

1. Data Prep: We cleaned the data, converted currencies to USD, calculated average subscription prices, and estimated revenue based on Substack's own "Paid Rank" tiers (e.g., "Thousands of paid subscribers" ~ 1000). We focused only on non-podcast newsletter posts.

2. Substack Revenue Model: We built a linear regression model predicting the (log-transformed and standardized) lower-bound estimated annual revenue. Predictors included average price, publication age (observation period), average time between posts, variance in time between posts, percentage of paid posts, average word counts (imputed where necessary), average description length, and category.

3. Post Likes Model: We built another linear regression model, this time predicting the (log-transformed) number of likes (reactions) on a post. Predictors included word count, description length, paid status, category, whether it was the *first* post, and crucially, the moving average of likes from the previous 10 posts.

We used standard statistical techniques, including log transformations to handle skewed data (like revenue and likes) and imputation for missing values. The goal wasn't perfect prediction but identifying significant drivers.

Results: What Drives Substack Success?

Predicting Which Substacks Earn More

Our model looking at publication-level revenue (Adjusted R-squared: 0.314, r=0.56, meaning it explains about 31.4% of the variance) revealed several significant factors:

Key Takeaways for Substacks:

• Charge More: Price remains a powerful lever. However, it should be noted that by default estimated revenue is a function of price. So this correlation may exist regardless.

• Post Often: Reducing the average time between posts (posting more frequently) still shows a strong positive association with revenue.

• Paid Percentage: The positive correlation holds – more paid posts generally link to higher revenue in the model, though the visual plots (below) still suggest a potential curve peaking around 50-60%.

• Word Count: Longer posts (both free and paid, after log transformations) still show a *statistically significant* positive correlation with revenue, but the effect sizes are smaller than before. Frequency likely remains more impactful than length alone.

• Consistency? Maybe Not: The slight positive correlation for *more* variance in posting intervals persists. Frequency seems to matter more than rigid timing.

Category Matters Too:

When we included categories in the model, the relative revenue potential (compared to the Arts & Entertainment baseline) showed this pattern:

Culture and US Politics remain top categories for revenue potential in this model, while Fiction and Philosophy show significantly lower potential.

Visualizing the Trends:

The visual patterns remain largely the same:

*Posting Interval vs. Revenue:* Shorter intervals trend higher.

*Average Price vs. Revenue:* Strong positive correlation.

*Substack Age vs. Revenue:* Older Substacks tend slightly higher.

*Percent Paid Posts vs. Revenue:* Linear trend up, but smoothed curve suggests a peak.

*Word Count vs. Revenue:* Positive trend, now reflecting the corrected word count scale.

Top Earning Substacks (Based on Estimates):

(If you want to copy+paste the links, visit the table on my site)

Predicting Post Popularity (Likes)

Our second model looked at factors predicting the number of likes (log-transformed) a post receives. This model was incredibly predictive (R-squared: 0.859)!

Key Takeaways for Posts:

• Momentum is Everything: The single biggest predictor by far is the moving average of likes on the previous 10 posts (`MA_10_posts`). Success breeds success. If your recent posts did well, your next one likely will too. This explains ~86% of the variance alone!

• Nail Your First Post: The `first_post` coefficient is enormous. Your very first post gets a massive visibility boost (algorithmic or otherwise). Don't waste it!

• Paid Wall Hurts Likes: Unsurprisingly, putting a post behind a paywall significantly reduces its like count. This is the trade-off for monetization.

• Length & Description: Longer posts get slightly *more* likes, while posts with slightly *shorter* descriptions do better. Keep the summary punchy?

• Category Effects: Comics and Health Politics posts tend to get more likes than average, while categories like Business, US Politics, and Technology get fewer, holding other factors constant.

Visualizing Post Popularity:

*Moving Average vs. Actual Likes:* The relationship is incredibly tight. Past performance is the best predictor of future performance.

*Model Predictions vs. Actual Likes:* Our model tracks actual likes very well, especially given the dominance of the moving average predictor.

Most Liked Posts (Raw Counts):

Best Performing Posts (Relative to Model):

These might represent posts with particularly viral topics, exceptional writing, or perhaps successful off-platform promotion.

Conclusion: Key Strategies for Substack Growth

Synthesizing the updated model results and observations:

1. Post Frequently: Still appears highly beneficial for revenue. Volume likely trumps perfectionism.

2. Build Momentum: Crucial for post visibility, given the massive impact of past likes.

3. Optimize Your First Post: Leverage that initial algorithmic (?) boost.

4. Price Strategically: Higher prices strongly correlate with higher revenue potential. Note the huge price difference between `nextplayinvesting` and `heathercoxrichardson` despite similar estimated revenue tiers – audience size and willingness to pay interact complexly.

5. Balance Free vs. Paid: The ~50% paid post mark still looks like a reasonable target based on visual inspection of the plots, despite the linear model showing a positive coefficient overall.

6. Consider Your Category: Significant differences in revenue potential exist between categories.

7. Word Count Matters (a little): Longer posts have a small, positive association with revenue, but don't sacrifice frequency for extreme length.

Ultimately, data provides patterns, not guarantees. Quality content and audience connection are paramount. However, understanding these underlying dynamics can help you make more informed decisions as you navigate the Substack landscape. Good luck!

• Global IQ is declining at 1.1 points per decade, with 35.5% of this decline attributable to dysgenics (fertility differences between individuals)

• The high-IQ working age population peaked between 1990-2040 and is now in decline; those with IQ ≥ 131 will decrease by 73.4% by 2100 (relative to 2025)

• By 2100, nearly 40% of the world's working-age population will have IQ <70, up from levels comparable to the 115-130 IQ bracket in the 1950s

• The +2SD IQ threshold (top 2.3% of population) will drop from 128 to 116 by 2100

• Global innovation capacity will halve by 2100, resulting in a loss of approximately 18 years of innovation potential this century

Introduction

Talented, industrious, and intelligent individuals are vital to humanity's progress. Their high productivity drives economic growth while their ingenuity advances science and innovation, uplifting everyone.

This relationship extends beyond individuals to nations as well. A country's National IQ (NIQ) serves as a powerful predictor of its scientific output, economic performance, and overall standard of living.

We can observe this connection in the relationship between the Innovation Index—a composite measure combining living Wikipedia figures in Discovery/Science, Nature Index publications, patents, and academic papers—and NIQ.

This pattern is further reinforced by the strong correlation between NIQ and the Socioeconomic Development Index, which integrates multiple indicators including mortality rates, infrastructure development, pollution levels, safety metrics, economic performance, educational attainment, and technological advancement.

If talented, intelligent individuals and nations are indeed crucial for development and innovation, then the declining numbers of such people should be cause for concern.

And declining they are.

Female fertility rates have been falling worldwide, with most countries now below replacement level—the threshold at which women have too few children to maintain population stability.

The issue isn't merely that overall fertility is declining, leading to shrinking populations. More troublingly, NIQ strongly negatively predicts fertility rates. Simply put: smarter countries have fewer babies. Consequently, the most developed nations with the highest concentrations of talent have experienced sub-replacement fertility for decades, with rates continuing to fall further.

The situation worsens when we consider individual-level data. Globally, individuals with higher education and income consistently produce fewer offspring than their peers, while those with lower cognitive ability tend to have more children.

Recent genomic studies from both the United States and United Kingdom confirm this pattern, showing that genetic variants associated with education and income are diminishing in frequency across generations.

These trends will likely have profound consequences for human economic development and wellbeing, amplifying the already concerning effects of declining fertility rates worldwide.

Innovation as Uncertainty

As concerning as these trends may be, they could potentially become irrelevant for an unexpected reason: artificial intelligence.

If we develop truly advanced AGI systems in the near future—systems capable of accelerating scientific discovery, transforming economic productivity, solving major health challenges including aging, or even posing existential risks—our future will become radically unpredictable.

Such a transformation would likely render concerns about gradual genetic selection against intelligence or declining fertility across generations largely moot. Our world would instead be continuously reshaped by exponentially compounding innovation generated by networks of intelligent machines operating globally.

This wouldn't be the first time technological breakthroughs have helped humanity avoid predicted disasters. Peak oil concerns dominated discussions in the late 2000s, but innovations in extraction technology and new reserve discoveries effectively navigated around the problem. Similarly with climate change, the rapidly falling costs of renewable energy and batteries are now driving down emissions in ways that seemed unlikely just a decade ago.

However, the singularity remains theoretical, and unlike peak oil or climate change, dysgenics and population decline directly affect the very engine of solutions: innovation itself. This creates a potentially dangerous feedback loop that, if not disrupted by transformative AI within the next few decades, could lead to a new dark age for humanity.

To understand this dynamic better, we should conceptualize innovation as fundamentally representing uncertainty about the future.

Forty years ago, no one could have accurately predicted how the internet and smartphones would transform virtually every aspect of modern life—from work and education to commerce, banking, socializing, and dating. It is sometimes said that genius hits the target that nobody sees. Likewise, if the world today is

the accumulation of the efforts of genius, the world is a series of changes nobody foresaw.

This principle works bidirectionally: just because transformative AI appears imminent doesn't guarantee it will materialize as expected, or at all. Dysgenics and declining fertility could continue to exert meaningful influence regardless.

Therefore, if innovation represents uncertainty, and if dysgenics and low fertility reduce innovative capacity, we face a self-reinforcing cycle: declining fertility and dysgenics reduce innovation, which in turn reduces uncertainty about our future trajectory. Greater certainty about continued population decline and dysgenics then further entrenches these trends.

This feedback mechanism is what makes declining birth rates and dysgenics truly concerning. Once innovation falls to half its previous level, the probability of developing transformative technologies capable of breaking this cycle diminishes accordingly.

If this loop remains unbroken, humanity could eventually enter a prolonged period of stagnation. With this sobering introduction, let's examine the data to estimate how much time remains before this cycle becomes difficult to escape.

To answer this question, we must explore the literature on dysgenics and fertility, integrating the most reliable quantitative data into a comprehensive forecast.

We'll then develop an accurate measure of innovation, use NIQ to predict it, and model the effects of demographic changes through linear analysis.

Our forecasting approach focuses primarily on global IQ trends both within and between countries. While intelligence is just one dimension of human capital, its effects are well-studied and readily interpretable, making it an ideal focus for our projections.

The Numbers: Dysgenics

As alluded to, dysgenics extends beyond IQ. Within human biodiversity research, the phenomenon is often conceptualized and measured in terms of intelligence, largely because IQ provides a convenient and interpretable metric. However, the dysgenic effect encompasses a broader range of heritable traits related to human capital.

Dysgenics in Genomics

The genomics literature offers compelling evidence for this broader dysgenic pattern.

A recent US study, Natural Selection Across Three Generations of Americans (Hugh-Jones & Edwards, 2024), analyzed data from the "Health and Retirement Study," which includes several thousand genotyped Americans born between 1920-1960. The researchers calculated each participant's relative lifetime reproductive success (RLRS) and used polygenic scores—genetic predictors of various traits—to identify which genetic profiles were associated with higher reproductive rates.

Their analysis spanned three generations by examining not only the participants' own reproductive outcomes but also their number of siblings (previous generation) and grandchildren (next generation).

The results revealed a consistent pattern, with the x-axis below showing effect sizes on reproductive success:

The data demonstrates that genetic variants associated with depression, lower educational attainment, poorer health, and lower cognitive ability predict greater reproductive success. Notably, ADHD shows the strongest positive selection, while educational attainment is the second most strongly selected against.

Given the known negative relationship between ADHD and educational achievement, the researchers investigated whether a trait's relationship to education might predict its selection pattern. Indeed, they found an extraordinarily strong negative correlation (r=-0.829) between a trait's association with education and its relationship to reproductive success:

This study built upon earlier research. Human Capital Mediates Natural Selection in Contemporary Humans (Hugh-Jones & Abdellaoui, 2022) conducted similar analyses using the UK Biobank dataset and likewise found that genetic variants associated with education and earnings were being selected against:

This second paper also mapped selection patterns across a wide range of traits, from those most negatively selected against to those most positively selected for. The consistency across traits is striking:

These findings demonstrate a robust pattern across both US and UK populations: genetic variants associated with depressive disorders are being selected for, while those linked to better health, higher cognitive ability, and greater educational attainment are being selected against.

Quantifying IQ Decline

How can we translate these selection effects into concrete estimates of IQ change over time? What do these "effect sizes" mean in practical terms?

A significant challenge in this analysis is the limited predictive power of current polygenic scores. These scores remain noisy predictors of actual phenotypic traits—not because genetics or heritability poorly predict traits, but because our genetic models are still developing.

This gap between known genetic influences and observed heritability from twin studies is known as "missing heritability." While this limitation will likely be resolved as genomic science advances, it currently requires adjustment in our calculations.

We can theoretically correct for this error by multiplying effect sizes by the square root of the ratio between trait heritability (h²) and polygenic score accuracy (r²). The first paper we discussed attempts this adjustment, yielding more reasonable effect estimates:

With these adjusted values, we can estimate the rate of IQ decline. The coefficient for "Cognitive Performance" on RLRS is approximately -0.06 using twin heritability. Let's convert this into a more interpretable IQ points per decade:

1. Start with the genetic selection coefficient: -0.06 standard deviations per generation

2. Convert to phenotypic change by multiplying by IQ heritability (h²):

• Using h² = 0.8 (80% heritability)

• -0.06 × 0.8 = -0.048 standard deviations

3. Convert from standard deviations to IQ points:

• IQ has a standard deviation of 15 points

• -0.048 × 15 = -0.72 IQ points per generation

4. Convert from per generation to per decade:

• Mean age of conception (MAC) in USA ≈ 31 years (UN data, 2023)

• Per decade rate = (10 years ÷ MAC) × per generation rate

• (10 ÷ 31) × -0.72 = -0.23 IQ points per decade

This calculation yields an estimated decline of -0.23 IQ points per decade for the USA.

How does this genomics-derived estimate compare with direct measurements?

The International meta-analysis of differential fertility for intelligence (Jensen et al., 2023), which directly measures IQ dysgenics rather than using genomic methods, found IQ declines of 0.1 to 0.38 points per decade across the Anglosphere, with -0.38 for the USA specifically:

Our genomics-based calculation (-0.23) falls remarkably close to these directly measured estimates!

While these calculations should be considered approximations rather than precise predictions, their alignment with measured values suggests they capture the general magnitude of the effect. They provide a reasonable guide to what traits are being selected for or against, and at what relative rates.

Selection Against IQ and SES

Moving from genotypic to phenotypic selection, we find that dysgenic effects vary significantly by region and socioeconomic context. Selection against intelligence appears stronger in less developed countries:

This pattern is mirrored in research on socioeconomic status (SES). The paper Fertility trends by social status (Skirbekk, 2008) systematically analyzed measurements of SES—including education, occupational status, and income/wealth—and compared reproductive outcomes between high and low status individuals across different time periods and regions.

The findings revealed not only that selection for SES has turned negative in recent decades, but also that this negative selection is more pronounced in developing regions:

This regional pattern is further reinforced by individual-level data from the genomic studies discussed earlier. Dysgenic effects were stronger among younger mothers and those with lower income and education:

This creates a striking parallel: the same factors—lower income and education—predict stronger dysgenic effects at both individual and national levels. While this could be coincidental, the consistency of the pattern across different scales and methodologies suggests a robust relationship.

To summarize, substantial evidence indicates selection against traits associated with human capital (education, SES, IQ) within populations. This selection appears particularly strong in lower-income countries and among lower-income individuals. While the magnitude of this dysgenic effect remains modest—typically fractions of an IQ point per decade—its cumulative impact over multiple generations could become significant.

The Numbers: Fertility, Population Projections, Aging Populations and Migration

While selection effects tell us about relative changes in trait distributions, absolute population changes are equally crucial for understanding the future of human capital. Selection is inherently relative—it describes how traits shift on average regardless of total population size. For our model to be comprehensive, we need to consider not just average IQ by country, but the actual number of individuals at each IQ level.

This distinction matters because selection processes as described earlier are relatively slow. In developed countries, it would take centuries for selection alone to shift average IQ by a full standard deviation.

If global fertility remained above replacement level, the absolute number of high-IQ individuals might still increase despite dysgenic trends, potentially offsetting concerns about declining innovation. However, as we've established, fertility rates have fallen below replacement level in most developed countries—precisely those with the highest average IQ.

The Inaccuracy of UN Forecasting

To project these population dynamics accurately, we need reliable population forecasts. The United Nations publishes updated projections every two years, but their accuracy has been questionable. For a more detailed analysis of UN forecasting methods and alternative approaches, see my previous analysis.

The limitations of UN projections are evident in historical comparisons, as shown below:

The UN has consistently underestimated the pace of fertility decline, particularly in rapidly developing regions. This pattern is clearly illustrated in recent data from the Financial Times, showing how fertility rates in South American countries have fallen far below UN projections:

Given this systematic bias, we can reasonably conclude that standard UN population projections are likely too optimistic. To compensate for this, our analysis uses the UN's "Low" fertility variant, which has historically tracked closer to actual outcomes.

Aging Population Dynamics

Population size alone doesn't tell the full story—age distribution is equally important, especially when considering innovation potential and economic output. Children contribute minimally to current innovation, while elderly individuals have typically produced their most significant contributions earlier in life.

The UN's methodology for projecting life expectancy introduces additional concerns. Their models assume linear increases in longevity, even for countries already approaching or exceeding 90 years of life expectancy:

This assumption distorts projections for countries with the combination of very low fertility and high life expectancy—a common pattern given the two are correlated. Such projections can create unrealistic scenarios where countries maintain large elderly populations despite minimal reproduction.

As emphasized in our introduction, our goal extends beyond simply documenting dysgenic trends—we aim to understand how these patterns, combined with demographic shifts, will affect the global pool of human capital and its capacity for innovation.

Migration Effects

Migration adds another layer of complexity to population projections. While migration doesn't alter an individual's genetic potential, it can significantly affect where and how that potential is realized.

Our model makes a simplifying assumption that focuses on genetic factors. This is admittedly an oversimplification—development clearly influences innovative output, as demonstrated by China's dramatic rise in scientific productivity. Talented individuals from developing countries likely realize more of their potential after moving to environments with better research infrastructure, education systems, and funding opportunities.

However, research also indicates that migrants and their descendants often retain characteristics associated with their ancestral populations. In many cases, especially in Europe, migrants are a net fiscal negative.

Since this analysis focuses primarily on global trends rather than migration specifically, we'll acknowledge these complexities without exploring them in depth.

Method

Our approach integrates multiple data sources to create a comprehensive model of global IQ distribution and innovation potential over time. The high-level overview of our methodology is as follows:

1. For each country, we compile:

• National IQ estimates

• Dysgenics estimates on NIQ per year

• Population projections by age, adjusting for migration and low fertility

• Innovation Index values

2. Using this data, we calculate for each country and year:

• National IQ by year (dysgenics-adjusted)

• Innovation Index, adjusted for the effect of dysgenics

• Working age population

3. Finally, we use population-weighted averages to determine for each year:

• The combined effect of dysgenics and declining population on innovation

• Average global IQ

• Additional demographic statistics

The following sections explain each component in detail, including data sources and calculation methods.

Dysgenics and NIQ

We obtained National Dysgenics rates from Jensen et al. (2023) and National IQ values from National IQs and Socioeconomic Development (Jensen & Kirkegaard, 2024).

For many countries, direct dysgenics measurements were unavailable. We addressed this gap by imputing missing values based on NIQ, which provides reasonable accuracy (r = 0.539, p<0.001).

To convert generational IQ decline rates to annual rates, we used the UN's data on mean age of conception by country. This conversion is necessary because dysgenics literature typically reports effects per generation, while our model operates on a yearly timescale.

Population Projections

Creating accurate population projections required addressing several limitations in standard UN data. The UN's modal "Medium" scenario consistently overestimates future populations, as demonstrated by historical comparisons.

To account for this bias, we used the UN's "Low" fertility variant. However, this introduced a complication regarding migration effects. The UN provides "Zero Migration" projections, but these are based on the "Medium" fertility scenario, not the "Low" variant we needed.

To resolve this issue, we developed a method to create "Low Fertility, Zero Migration" estimates:

1. We first confirmed that the UN uses identical migration figures for both Low and Medium variants:

2. We then calculated the difference between the Medium and Zero Migration variants by age group.

3. Since the migration figures are identical between variants, we could apply this same difference to the Low variant to obtain our desired "Low Fertility, Zero Migration" estimates.

This approach preserved the age-specific migration patterns in the UN data while adjusting for the more realistic fertility projections in the Low variant.

Innovation Index Construction

To measure innovation potential, we created a composite index from four key metrics:

1. Notable scientists from A cross-verified database of notable people, 3500BC-2018AD

• We extracted the number of living figures in Discovery/Science by country as of 2020

• This database draws from Wikipedia across all languages via Wikidata, not just English Wikipedia

2. Nature Index share

• Publications in Nature journals, adjusted for author count

• If 10 authors from 10 different countries collaborate on one paper, each country receives a score of 1/10

3. Scientific publication count

• Total papers published, obtained from Our World In Data

• We converted per-million figures to absolute counts by multiplying by 2020 population

4. Patent application count

• Total patent applications, obtained from Our World In Data

• Similarly converted from per-million to absolute counts

We transformed these metrics into log-count values rather than per-capita measures, as log-counts generally exhibit better statistical properties. Zero values were treated as missing data rather than true zeros, recognizing that small or low-NIQ countries might not register in these metrics due to chance rather than a complete absence of innovation.

We imputed missing values using R's MICE package and then extracted the first principal component to create our Innovation Index. The resulting index showed strong correlations with all input metrics and with NIQ:

An important characteristic of these innovation metrics is their exponential relationship with NIQ. As NIQ increases, innovation output increases exponentially rather than linearly. This means low-NIQ countries produce disproportionately little innovation, while high-NIQ countries produce disproportionately more. Our log-based index accounts for this non-linearity.

For calculating actual innovative contribution by country, we applied the inverse log operation to convert back to the original scale.

Results

IQ

Let's examine our findings, beginning with a visualization of how the global IQ distribution changes over time:

Our analysis reveals that the average IQ for the working-age population is declining at approximately 1.1 IQ points per decade. Furthermore, dysgenics accounts for 35.5% of this decline, with the remainder attributable to demographic shifts—specifically, higher-fertility populations with lower average IQ becoming a larger proportion of humanity. Among toddlers, the projected decline is even more dramatic, with average IQ falling from about 91 to 75 between 1950 and 2100.

How do these changes affect the actual population distribution across different IQ ranges? Despite shifts in averages, are we seeing meaningful reductions in the number of high-IQ individuals?

This question is challenging to visualize directly because the population with IQ >130 is so small that it's barely visible when plotted alongside other IQ ranges. To address this, we've used a logarithmic scale for the y-axis. The data shows that the working-age population with IQ >115 has already begun declining:

Perhaps most alarming is the changing proportion of the world population at different IQ levels. The proportion with IQ <70, which was once roughly equivalent to the 115-130 bracket, has increased dramatically since the 1950s. By 2100, this group is projected to comprise nearly 40% of the world's working-age population:

What about the overall distribution of intelligence? While the standard deviation of global IQ remains relatively stable, the shape of the distribution is changing. The symmetry of the normal curve shifts, with the +2SD threshold moving further from the median while the -2SD threshold moves closer, creating a slight right-tail skew in the global IQ distribution:

This shift also means that what constitutes an exceptional or below-average IQ is changing over time. The thresholds for various standard deviations are declining substantially. For example, the +2SD threshold (representing the top 2.3% of the population) drops from approximately 128 to 116:

Innovation

Given these projected changes in global IQ distribution, what are the implications for innovation? Our model suggests that the combination of a smaller working-age population and dysgenics will approximately halve global innovation by 2100 relative to 2023 levels:

While this decline is certainly undesirable, I must acknowledge that it appears survivable. To quantify the impact, we can calculate the cumulative innovation potential from 2025 to 2100 as "2023-equivalent years"—essentially asking how many years of innovation at 2023 levels we can expect during this 75-year period.

For the scenario without dysgenics, we project approximately 68 years of 2023-level innovation during the 75-year period. With dysgenics factored in, this drops to about 57 years—a loss of roughly 18 years of innovation potential.

This means that even in the dysgenic scenario, humanity retains substantial innovative capacity through the end of the century, though at a reduced level compared to what might have been possible.

Discussion

The projected loss of 18 years of innovation potential this century is relatively modest when viewed in historical context. The remaining 57 years of 2023-equivalent innovation represents substantial creative capacity—equivalent to all progress since 1968 (2025 - 57). That humanity could maintain this level of output despite aging populations and diminishing human capital is noteworthy.

However, our projections may actually underestimate the full impact of dysgenics. As our analysis has shown, selection is occurring against multiple traits associated with human capital, not just IQ. Even if we adjust our estimates to account for this broader effect—perhaps increasing the innovation loss to 25 years out of 75—there still appears to be sufficient time for transformative technological developments.

Consider the remarkable advances in AI and genetics over the past half-century, or even the extraordinary progress in AI capabilities just within the last five years. The pace of technological change suggests that even with reduced human capital, breakthrough innovations remain possible.

Beyond technological acceleration, several demographic and developmental factors could mitigate the projected decline in innovation:

China's emergence as a scientific powerhouse represents one such counterbalance. Its share in the Nature Index has grown by approximately 15% annually for nearly a decade, resulting in a dramatic shift in the global innovation landscape. China now produces about 14% more papers in Nature than the entire United States—a remarkable reversal from just a few years ago when its output was less than half that of the US.

This transformation is particularly significant given the strong correlation (r=0.95) between the Nature Index and our broader innovation measure, suggesting that China's rise represents a genuine increase in innovative capacity rather than merely bibliometric inflation.

The Chinese example suggests that economic development can unlock latent innovative potential in populations. This pattern could potentially extend to other developing nations as their GDP per capita increases. While human capital constraints will present challenges, countries with large populations like India—projected to approach 2 billion people this century—could collectively contribute innovation equivalent to multiple European countries combined if they follow a similar developmental trajectory.

This developmental effect applies not only to countries but to individuals as well. When talented people from lower-resource environments migrate to contexts with better research infrastructure and opportunities, they may realize more of their innovative potential. This mobility effect could partially offset the impact of dysgenics within populations.

Taken together, these factors—technological acceleration, the rise of developing nations, and human capital mobility—provide reasonable grounds for cautious optimism despite the concerning trends in global IQ and fertility.

Conclusion

In conclusion, it's likely that dysgenics and declining fertility will have a meaningful impact on global innovation and human capital in the coming decades. Our analysis projects a halving of innovation by 2100 relative to 2023 levels, with a loss of approximately 18 years of innovation potential this century.

However, this forecast should be viewed with cautious optimism. The projected 57 years of 2023-equivalent innovation remaining this century still represents substantial progress. For context, the last 57 years have transformed our world beyond recognition through computing, the internet, mobile technology, and the beginnings of AI.

Several factors could mitigate or even reverse these trends:

First, the rise of China and potentially other developing nations demonstrates that economic development can dramatically increase innovative output relative to developed countries. As more countries develop robust infrastructures pertaining to innovation, they may unlock previously untapped human potential.

Second, migration from lower to higher-resource environments may allow talented individuals to better realize their capabilities, partially offsetting the effects of dysgenics within populations.

Third, and perhaps most significantly, technological acceleration itself could render these projections obsolete. AI systems are already enhancing human productivity across multiple domains. If this trend continues, even a smaller pool of high-IQ individuals might achieve more than larger populations could in previous eras.

The feedback loop between dysgenics, declining fertility, and reduced innovation remains a legitimate concern. However, the window for technological breakthroughs that could break this cycle appears sufficiently wide. With 57 years of substantial innovation ahead, humanity has time to develop solutions to these demographic challenges—whether through advanced AI, genetic technologies, or social innovations that reverse fertility trends among high-capability individuals.

The future remains uncertain, but that uncertainty itself is cause for hope. Just as no one in 1968 could have predicted our current technological landscape, the innovations of the coming decades may well transform our understanding of human potential and productivity in ways we cannot yet imagine.

Hugh-Jones D, Abdellaoui A. Human Capital Mediates Natural Selection in Contemporary Humans. Behav Genet. 2022 Sep;52(4-5):205-234. doi: 10.1007/s10519-022-10107-w. Epub 2022 Jul 6. Erratum in: Behav Genet. 2022 Sep;52(4-5):235. doi: 10.1007/s10519-022-10110-1. PMID: 35790706; PMCID: PMC9463317.

Hugh-Jones D, Edwards T. Natural Selection Across Three Generations of Americans. Behav Genet. 2024 Sep;54(5):405-415. doi: 10.1007/s10519-024-10189-8. Epub 2024 Jul 11. PMID: 38990442.

Jensen, S., Pesta, B. J., Francis, G., & Kirkegaard, E. O. W. (2023). International meta-analysis of differential fertility for intelligence. Open Psych. https://osf.io/2fb4p/

Skirbekk, Vegard. (2008). Fertility Trends by Social Status. Demographic Research. 18. 145-180. 10.4054/DemRes.2008.18.5.

Jensen, S., & Kirkegaard, E. O. W. (2024). National IQs and socioeconomic development. PsyArXiv. https://doi.org/10.31234/osf.io/bx86g

TL;DR

• Rich dataset of 1.2 million criminals.

• 160k+ faces analyzed for CNN.

• Convicted pedophiles are most likely to be older, white, overweight men.

• Our model achieved 69% accuracy in classifying pedophiles from facial features alone.

Introduction

When ImageNet debuted in 2009, it revolutionized computer vision with its dataset of over a million images spanning more than 1,000 object categories. The first ImageNet competition in 2010 saw a winning accuracy of just 52.9%, but everything changed two years later when AlexNet achieved a breakthrough 63.3% accuracy—outperforming the runner-up by a staggering 10%.

Although modest by today's standards, this achievement ignited the neural network "deep learning" revolution, transforming neural networks from academic curiosities into the foundation of modern AI. In the following years, innovations accelerated rapidly. By 2017, models surpassed 90% accuracy on the 1,000-class challenge, with top-5 accuracy exceeding 95%—better than human performance.

Remarkably, 2017 was only seven years ago, yet in the rapidly evolving field of machine learning, it feels like ancient history. Today's models vastly outperform anything from that era.

To the point.

Modern AI research has largely pivoted toward developing "everything" models—systems that understand and generate all forms of media (audio, video, text, images) and can solve problems across countless domains. The pursuit of artificial general intelligence (AGI) dominates the field.

Few serious ML scientists now focus on narrow, specific tasks. This is partly because the oxygen has been sucked out of the room with focus on LLMs. However, it's also because there's nothing novel in doing the next run of the mill CNN anymore. Predicting attributes such as age, gender, race, body mass index (BMI), facial expressions have all been solved. Or if they haven't been solved, it's only because a dataset hasn't been assembled to do so. Lastly, the remaining low hanging fruit could be considered controversial - predicting political beliefs, religiosity, personality or criminality.

However, some have still pursued this line of research in the 2020s. Hashemi and Hall (2020) published research demonstrating that convolutional neural networks could distinguish between "criminal" and "non-criminal" facial images with a reported accuracy of 97% on their test set. While this paper was later retracted for ethical concerns rather than methodological flaws, it highlighted the potential for facial analysis to extend beyond physical attributes into behavior prediction. Similarly, Kosinski (2021) demonstrated that facial recognition technology could predict political orientation from naturalistic facial images with 72% accuracy, significantly outperforming human judges who achieved only 55% accuracy. These models maintained substantial predictive power even when controlling for demographic variables such as age, gender, and ethnicity.

We'll continue that trend: predicting pedophilic behavior based solely on facial features, bringing levity to an otherwise serious crime.

Dataset

Criminal Mugshots and Offenses Collection

Without readily available public datasets for this task, I turned to my existing database of approximately 2 million mugshots collected from previous scraping projects. This data comes directly from various state Department of Corrections (Oklahoma as an example) websites across the US.

The dataset contains multiple mugshots for some individuals, and not all entries include offense information. After cleaning and preprocessing, I retained 1.2 million unique criminal mugshots with documented offenses, allowing me to differentiate between pedophiles and other criminals.

Preliminary Linear Modeling

The full dataset often includes demographic and physical attributes: race, gender, hair color, eye color, BMI, and more. To begin my investigation, I ran a logistic regression model to identify which basic characteristics correlate with pedophilia.

For reference in interpreting the results:

• Race comparisons use White as the baseline

• Hair color comparisons use Black as the baseline

• Eye color comparisons use Brown as the baseline

The p-values reveal that most characteristics show statistically significant correlations:

Significance codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ’ ’ 1

For easier interpretation, here are the odds ratios for statistically significant attributes:

Translating these odds ratios into plain English, our data suggests that convicted pedophiles tend to be:

• Male (over 10 times more likely than females)

• Older (indicated by gray hair, baldness, and gray eyes)

• Overweight (higher BMI correlates with increased likelihood)

• White (all other racial groups show lower odds ratios)

While this linear model's overall predictive accuracy barely outperformed random guessing, our enormous sample size allowed us to identify statistically significant demographic patterns. The consistent associations with age markers (gray hair, baldness, gray eyes) lend credibility to these findings. Personally, I think these patterns align with common stereotypes about pedophile offenders.

Now, let's move to our more sophisticated approach using convolutional neural networks.

PedoAI CNN Methodology

Creating this AI involved four key steps:

1. Classifying Pedophilia-Related Offenses with LLMs

Criminal offenses appear in widely varying formats across our dataset:

• "Poss W Purp Del Cont Sub LSD => 80 DU < 160 DU"

• "Theft of leased/rented property =>,000"

• "FACILITATION MARIJUANA VIOLATION"

With approximately 33,000 distinct offense descriptions among 1.2 million criminals, manual classification would be prohibitively time-consuming. I leveraged Gemini Flash 2.0 to identify pedophilia-related offenses, completing the task for about $1.

2. Extracting and Standardizing Facial Images

For our model to focus purely on facial features—not clothing, background, or other variables—we needed to isolate and standardize faces from the mugshots. This process involved:

1. Using multiple face detection models from the DeepFace library (MTCNN, YOLOv8, RetinaFace) with fallback options if primary detection failed

2. Resizing all extracted faces to 160 × 224 pixels while preserving aspect ratio using black padding

I determined these dimensions after analyzing 10,000 sample faces:

The 160 × 224 pixel dimensions were chosen because:

• Both numbers are divisible by 32 (computationally efficient)

• The width-to-height ratio (1.4) closely matches the natural proportions found in our dataset (1.38)

• 224 pixels is a standard height for facial recognition models

Here's a visual example of our extraction process:

3. Data Selection for Training

While our 1.2 million image dataset provides robust training material, training on the full set would require excessive computational resources, especially during hyperparameter optimization (which requires multiple training runs).

I struck a balance by using 100,000 pedophile images and 100,000 non-pedophile images for training, with separate validation and test sets of approximately 20,000 images each.

4. Training with Progressive Scaling

To understand how model performance scales with dataset size, I conducted 50 hyperparameter tuning runs at each of the following sample sizes: 5k, 10k, 20k, 40k, 80k, and the full 160k faces. This required training a total of 300 distinct models while maintaining consistent validation and test sets.

Results

Rather than presenting dry performance tables, I decided to test our model in a more engaging way. I connected PedoAI to a browser and let it play pedoguessr, a game where players guess which of two faces belongs to a convicted pedophile.

How did it perform? Unfortunately, not well in this context. Since my model's accuracy isn't exceptionally high, performance in this game largely came down to luck. Random guessing can achieve scores of 20-30, so I needed many trials to properly evaluate the AI's capabilities.

Taking advantage of a game exploit that allowed indefinite play (earning the username "Filthy Cheater!" the #1 spot on the leaderboard 😁), I let the AI complete 303 rounds. The results were disappointing: wrong in 160 out of 303 rounds—worse than random guessing.

Even more concerning, the model's confidence levels didn't correlate with its accuracy. In theory, when the AI sees a bigger difference in predicted probabilities between two images (e.g., 1% vs. 75% likelihood of being a pedophile), it should be more likely to get the answer right than when the probabilities are close (51% vs. 52%). However, statistical analysis showed no significant relationship:

Despite these disappointing game results, the model performed much better on our controlled test set. On a sample of 20,000 faces the model never saw during training, it achieved 68.88% accuracy—significantly better than random chance:

We also observed clear scaling benefits with larger training datasets:

Why did our model struggle with the pedoguessr game despite its reasonable performance on our test set? The discrepancy likely stems from different labeling criteria. The game creator may have used different standards to classify pedophiles, perhaps focusing on different offense types or severity levels than those in our training data. This highlights a key limitation: while our model can predict better than random chance in controlled settings, it lacks broad generalizability to other contexts.

Conclusions

The results of this study present a complex picture of AI's ability to predict pedophilic tendencies from facial features. While our model achieved nearly 69% accuracy on our test dataset, several important limitations must be acknowledged.

The study by Hashemi and Hall (2020) provides a cautionary tale about facial recognition models potentially overfitting to characteristics unrelated to facial morphology. Their model achieved a suspiciously high 97% accuracy while showing continual improvement as they added more convolutional layers, despite working with a relatively small dataset of just 10,000 images. This pattern of unending improvement with increased model complexity often signals overfitting rather than meaningful feature learning. In contrast, our approach used a much more modest architecture of only 4 layers despite having access to over 10 times their sample size. During our hyperparameter tuning across different sample sizes, we found that model performance did not universally improve with additional layers—instead, there was typically an optimal layer count beyond which performance plateaued or declined, suggesting a more honest learning process. Additionally, Hashemi and Hall's study drew criminal and non-criminal images from fundamentally different sources, potentially allowing their algorithm to detect irrelevant environmental differences such as:

• Standardized lighting conditions

• Institutional clothing

• Controlled facial expressions

• Demographic characteristics typical of the university students who comprised their non-criminal sample.

Our approach mitigates these concerns by using mugshots from the same database for both pedophiles and non-pedophile criminals, significantly reducing the risk of systematic differences in image characteristics between our comparison groups.

Nevertheless, our model's poor performance on the external "pedoguessr" dataset, where it performed worse than random chance, suggests limited generalizability beyond our specific dataset. As Kosinski (2021) noted in his work on political orientation prediction, even well-performing models can fail to transfer across contexts when they learn dataset-specific patterns rather than robust physiological markers.

The Bayesian prior issue presents another significant limitation. Our model was trained on a balanced dataset containing 50% pedophiles and 50% non-pedophile criminals. This equal distribution dramatically overrepresents the actual prevalence of pedophiles in the general population. In real-world applications, even a model with seemingly impressive accuracy metrics would generate an unacceptable number of false positives when applied to a population with a much lower base rate of the target condition. While this could be addressed through threshold adjustments, such tuning would inevitably reduce the model's ability to correctly identify actual pedophiles (reduction in false positives leads to an increase in false negatives through this method).

Perhaps most fundamentally, we may be attempting to make distinctions between groups that are topologically close in feature space. Both groups in our study consist of individuals convicted of crimes - pedophilia representing one specific criminal category. The physiological and demographic similarities between different types of criminals likely create substantial overlap that makes clean classification inherently difficult. Unlike distinguishing between broader categories (such as political orientation, as demonstrated by Kosinski, 2021), differentiating between subtypes of criminals may require signals too subtle for current computer vision techniques to reliably detect.

This post highlights the significant gap between laboratory performance and real-world applicability of facial analysis for predicting complex behavioral traits. While our results show some statistical relationship between facial features and criminal pedophilic behavior, the limitations in generalizability, base rate considerations, and potential dataset artifacts suggest caution is warranted.

Hashemi, M., & Hall, M. (2020). Criminal tendency detection from facial images and the gender bias effect. Journal of Big Data, 7(2), 1-16. https://doi.org/10.1186/s40537-019-0282-4

Kosinski, M. (2021). Facial recognition technology can expose political orientation from naturalistic facial images. Scientific Reports, 11(100), 1-7. https://doi.org/10.1038/s41598-020-79310-1

TL;DR

• AI massively improves job performance.

• AI improves job performance for less competent, junior developers more than it does senior developers.

• H1Bs, excessive CS graduates, and tech layoffs have created a glut of labor supply.

• The theoretical labor supply has increased even further, since AI has made programming accessible to a large population pool.

Introduction

Recently, there were debates over H1B visas. The debate's two camps were as follows:

• In support, Elon Musk, Indians and his followers. By increasing H1Bs, they argue this will expand the overall developer pool in quantity and quantity of quality and will thus keep US companies competitive with lower wages and greater talent pools.

• The opposition consisted of a bipartisan coalition, from leftists on Reddit to the dissident right on X. By abating H1Bs, they argue jobs will be kept for American developers and graduates, which are sufficient in quantity and quality to fulfill the needs of the US tech companies.

Instead of subsiding as 'current thing' debates usually do, it maintained intensity for weeks, eventually expanding in scope. Popular conservatives commentators began siding with the Elon faction arguing that young university educated men should suck it up. If their jobs are taken by H1Bs, then they should pursue careers at fast food chains instead.

The conversation sort of shifted away from tech employment and more towards generational inequality, focusing on whether life was harder for Zoomers over Boomers, etc. This is an oversimplification of course. Given the intensity and scope of the debate, many peripheral subjects were touched.

However, one insufficiently addressed subject - in my opinion - was the shifting landscape of tech employment pertaining to the impact of AIs and overall credential inflation (devaluing of education credentials at every level of educational attainment).

How are these tangentially related?

For AI, especially for developers, it may boost developer productivity by an amount comparable to over a standard deviation increase in IQ, based on typical IQ / job performance correlations.

Since the underlying complaint of the H1B opposition was that, given an increase in quantity of H1Bs, quality may necessarily fall since H1Bs would become less selective. H1B opponents are usually thinking of IQ when they think about quality. However, if the job performance gains from AI are large, then it doesn't matter if they're low IQ because AI can simply enhance their performance to where this simply doesn't matter.

Of course, H1B opponents would simply retaliate by saying that, if low IQ fellows can program with AIs, then why need H1Bs at all? One could simply use nationals of deficient talent - not foreigners - as they are still in sufficient supply.

This is especially true if we consider that there are more students graduating with degrees in computer science per capita.

But this got me thinking, if anyone can be a developer now, and there's a glut of workers on the market, then being a developer isn't very prestigious anymore, is it? All the market forces would be pushing downward on developer salaries, as there is now a legion of applicants with sufficient post-AI ability.

Now curious, I wanted to see some real data on the situation. How much does AI improve developer, and non-developer professional performance? What's the equivalent IQ x job performance effect size that we see? Are there any other nuances in the data?

AIs makes us stupid smart

After some digging and citation hopping, these are the main papers on productivity gains with AI. You can click on the header for the source. Unfortunately not all effects are reported in standard units.

Note: *** indicates statistical significance at p<0.01, * indicates significance at p<0.10. No * either means not significant or significance not reported.

Brynjolfsson et al. (2023)

*Task: Customer support agents at a Fortune 500 software firm handling customer inquiries through chat*

• 14% increase in customer support resolutions per hour overall***

• 34% increase in resolutions per hour for novice workers***

• 9% decrease in average customer chat duration***

• 1.3% increase in successful chat resolution rate*

Peng et al. (2023)

*Task: Professional programmers implementing an HTTP server in JavaScript*

• 55.8% reduction in time to complete server implementation task***

• 7% higher task success rate in completing implementation requirements

Gambacorta et al. (2024)

*Task: Software programmers at Ant Group working on regular coding tasks*

• 55% increase in lines of code produced overall***

• 67% increase in lines of code produced by junior staff***

• 11-18% of productivity gains directly attributable to LLM code output

Cui et al. (2024a)

*Task: Software developers at Microsoft, Accenture, and Fortune 100 company performing regular development work*

• 54.03% increase in completed pull requests at anonymous company

• 38.38% increase in code compilation attempts***

• 26.08% increase in completed pull requests overall***

• 13.55% increase in code commits

• 5.53% decrease in successful build rate

Yeverechyahu et al. (2024)

*Task: Open-source developers contributing to Python and R packages*

• 51% increase in code commits to repositories***

• 17.82% increase in new package version releases***

• 15.14% higher increase in maintenance-related commits compared to feature development***

Cui et al. (2024b)

*Task: Software developers at Microsoft and Accenture performing regular development work*

• 84-107% increase in successful code builds at Accenture***

• 12.92-21.83% increase in completed pull requests at Microsoft***

• 11.53% increase in lines of code changed***

• 7.51-8.69% increase in completed pull requests at Accenture***

McKinsey (2023)

*Task: Software developers performing various coding tasks including documentation, generation, and refactoring*

• 45-50% reduction in time spent on code documentation

• 35-45% reduction in time spent on new code generation

• 20-30% reduction in time spent on code refactoring

• <10% reduction in time spent on complex programming tasks

Vaithilingam et al. (2022)

*Task: Students and engineers completing Python programming assignments*

• ~1 minute faster task completion time

• Significantly higher helpfulness rating (6.16 vs 4.45 out of 10)***

Mozannar et al. (2024)

*Task: Software developers completing pre-selected coding tasks*

• 55.8% potential reduction in overall task completion time

• 76% of participants reported improved productivity (16/21 participants)

• 81% of participants reported faster task completion (17/21 participants)

Campero et al. (2022)

*Task: HTML "programmers" and non-programmers creating web pages*

• 27% improvement in task completion speed (using regression method)***

• 17% improvement in task completion speed (using ratio of means)***

Noy & Zhang (2023)

*Task: College-educated professionals completing occupation-specific writing tasks*

• 37% or 0.8 SD reduction in task completion time (from 27 to 17 minutes)***

• 0.45 standard deviation increase in output quality***

• 33% vs 18% adoption rate post-experiment***

• 0.40 standard deviation increase in job satisfaction***

• 0.20 standard deviation increase in self-efficacy*

Okay, so there are some pretty sizable gains. What was the method? What generation of AIs did they use, etc?

AIs makes us stupid, smart

Here's a rough table of the method, sample size, when the study was conducted, and the rough equivalent generation of AI that was employed, and the AI that was reported in use. As you can see there are many RCTs here that are well sampled.

What's astonishing here are the AIs used in the studies.

Early versions of co-pilot which are glorified autocompletes and GPT-3.5.

THESE AIs ARE AWFUL.

GPT-3.5 was unusable, it hallucinated continuously, couldn't do kindergarten maths, it had a memory of a few thousand words, and even then it only remembered either the beginning or the end of your conversation. There isn't a single AI in these studies that was even as advanced as GPT-4o.

But DESPITE all that, as reported above, the gains were enormous!

An interesting finding I found perusing the papers was the presence of an interaction effect between productivity gains and developer experience or competence.

Less experienced devs benefited more! See negative interaction field.

This negative interaction effect is theoretically devastating for developers, and any professional class apparently according to these papers. This is because it closes the gap between the incompetent and the competent, the low and high IQ, the inexperience and the experienced. Why would you hire a talented experienced senior dev when you can save 50% by going with a junior dev that can nearly get the same job done with AI? The same could be said about H1Bs. Employ foreigners that are beholden to you though a visa, pay them less, and get the same performance because they're using AI.

By how much does AI increase developer IQ?

Okay, but what's the ballpark equivalent IQ gain? This depends on three assumptions:

1. The correlation between IQ and job performance.

2. "Job performance" is analogous to "productivity gains" we've just mentioned.

3. Guessing from the various papers, which are heterogenous in how they measure productivity, the standard deviation gain in productivity or job performance.

We don't really have to assume point one, but it is fiercely debated - check out Meng Hu's great article. From reading his work, it seems that the correlation between IQ and productivity of about r=0.4.

Point two difficult to ascertain because, as it is even in job performance x IQ papers, measurements are heterogeneous. Not every paper is using the same method to measure job performance, and not every AI productivity paper is either. So we need to assume that the two are mostly analogous. I think this is likely true, but it is an assumption.

Lastly, and annoyingly, not every paper is reporting standardized gains. Completing a task 37% faster might sound impressive, but if it only equates to a performance gain of 0.2 SDs, then it's misleading. So we have to ballpark guess.

Given the above, I wouldn't take the below too seriously. However, given that the gains that were saw in the papers above were from low quality AIs, in a way these should be considered lower bound.

My personal guess:

• 95% CIs 0.1-0.6 SD improvement or 0.35 SD for seniors.

• 95% CIs 0.3-1 SD improvement or 0.65 SD for juniors.

Assuming IQ x job performance correlation of 0.4, a 0.35 and 0.65 SD gain, this is a 0.875 or 1.625 SD IQ gain equivalent. That's about ~13 or ~24 IQ points respectively.

Accessing the supplementary materials of this paper, it appears that "Programmers and software development professionals" have an IQ of about 111.2 (it’s in the supplementary materials).

So maybe a developer from a third world country with an IQ of ~90 can perform as well as a first world programmer post AI? That seems to be the implication here if we take the ~24 point gain at face value.

Conclusion, Discussion

Recent data suggests AIs have democratized programming by effectively adding >15 IQ points to developers' capabilities. The effect is most pronounced among less competent programmers, with multiple RCTs showing 30-70% productivity gains using even primitive AI models like GPT-3.5.

This reduced barrier to entry, combined with credential inflation (surge in CS graduates), theoretically should be negative for developer employment. Why hire senior devs when juniors with AI assistance can perform similarly at half the cost? AI doesn't just augment talent, it flattens the skill distribution.

The implications are stark. Developer salaries will likely continue declining as the market saturates with AI-augmented developers. The profession's prestige diminishes as programming becomes increasingly accessible.

Already, the job situation hasn't returned to pre-COVID levels.

The future may see programming transform from a high-skill profession to a commodity skill, with AI serving as the great equalizer. That is until AI replaces the occupation entirely.

Brynjolfsson, E., Li, D., & Raymond, L. (2023). Generative AI at work. arXiv. https://doi.org/10.48550/arXiv.2304.11771

Campero, A., Vaccaro, M., Song, J., Wen, H., Almaatouq, A., & Malone, T. W. (2022). A test for evaluating performance in human-computer systems. arXiv. https://doi.org/10.48550/arXiv.2206.12390

Cui, Z., Demirer, M., Jaffe, S., Musolff, L., Peng, S., & Salz, T. (2024a). The effects of generative AI on high skilled work: Evidence from three field experiments with software developers. SSRN. http://dx.doi.org/10.2139/ssrn.4945566

Cui, K., Demirer, M., Jaffe, S., Musolff, L., Peng, S., & Salz, T. (2024b). The productivity effects of generative AI: Evidence from a field experiment with GitHub Copilot. MIT Exploration of Generative AI. https://doi.org/10.21428/e4baedd9.3ad85f1c

Gambacorta, L., Qiu, H., Shan, S., & Rees, D. (2024). Generative AI and labour productivity: A field experiment on coding (BIS Working Paper No. 1208). Bank for International Settlements. https://www.bis.org/publ/work1208.htm

McKinsey Digital. (2023, June 27). Unleashing developer productivity with generative AI. McKinsey & Company. https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/unleashing-developer-productivity-with-generative-ai

Mozannar, H., Bansal, G., Fourney, A., & Horvitz, E. (2024). Reading between the lines: Modeling user behavior and costs in AI-assisted programming. arXiv. https://doi.org/10.48550/arXiv.2210.14306

Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187-192. https://doi.org/10.1126/science.adh2586

Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The impact of AI on developer productivity: Evidence from GitHub Copilot. arXiv. https://doi.org/10.48550/arXiv.2302.06590

Vaithilingam, P., Zhang, T., & Glassman, E. L. (2022). Expectation vs. experience: Evaluating the usability of code generation tools powered by large language models. In CHI Conference on Human Factors in Computing Systems Extended Abstracts (CHI EA '22). Association for Computing Machinery. https://doi.org/10.1145/3491101.3519665

Yeverechyahu, D., Mayya, R., & Oestreicher-Singer, G. (2024). The impact of large language models on open-source innovation: Evidence from GitHub Copilot. arXiv. https://doi.org/10.48550/arXiv.2409.08379

Hu, M. (2024, June 2). Controversy over the predictive validity of IQ on job performance. Substack. https://menghu.substack.com/p/controversy-over-the-predictive-validity-of-iq

Wolfram, T. (2023). (Not just) Intelligence stratifies the occupational hierarchy: Ranking 360 professions by IQ and non-cognitive traits. Intelligence, 98, Article 101755. https://doi.org/10.1016/j.intell.2023.101755

• Age of first sexual intercourse has declined since the sexual revolution. The effect is stronger in women than men.

• Total sexual partner count, or new partner/s in the past year, is either increasing or stagnant in most countries.

• Sexual frequency (how often people have sex) is declining.

• Declining marriage rates partially responsible for sex frequency decline.

**Consider reading on the main website uncorrelated and following my X/twitter**

Introduction

We know sexlessness is rising in adolescence based on the YRBSS, GSS and to a lesser extent the NSFG.

But there's a problem with all these datasets they're all American!!!

While the US data has received the most attention, particularly from incel communities, we shouldn't assume these trends are universal across all developed nations.

Naturally, as a result of this, many have generated causal theories attempting to explain rising youth sexlessness. These range from ideological conflicts between the genders, Chadopoly and dating apps (pushed by Incels), and some tech based ones with smart phones and social media (pushed by Jean Twenge), Jonathan Haidt).

What underpins most causal explanations is that they could be applied to any developed country; everyone has access to dating apps, smartphones, left/right political divides are almost everywhere, and there's decent evidence mental illness is on the rise internationally too.

Any theory attempting to explain a cause assumes that there is something to explain in the first place. If there is no increase, decrease, correlation, etc, then there's nothing for the cause explain.

So if we demonstrate the macro timeseries trends for sexlessness are null or contra the US internationally, we can at least increase skepticism of these explanations.

Sexlessness, Internationally

To this end, we will attempt to investigate the long running trends in sexlessness. Once again, we are interested in the key variables:

1. Age at first sex

2. Sexual frequency

3. Lifetime partner counts

Unfortunately, ~~all~~ most datasets from around the globe measuring sex analogous to their US counterparts are behind ethics committees.

So we have to rely on what's reported by academics in the literature specific to each dataset for every country.

United Kingdom

United Kingdom (NATSAL)

NATSAL is a well resourced, representative survey conducted decennially. NATSAL focuses on a broader age range, rather than adolescents.

Overall, 15 162 interviews were completed, with a response rate of 57.7% and a cooperation rate of 65.8%. The response rate for the boost sample of ages 16–34 years was 64.8%, only marginally lower than the 65.4% achieved for Natsal-2, which surveyed a similar age range (16–44). The data were weighted by age, gender and region to reduce possible bias. Comparisons with census data show the weighted sample to provide good representation on a range of respondent characteristics. The interview involved a combination of face-to-face and self-completion components, both carried out on computer

NATSAL has conducted five surveys between 1990-2024, with results from the latest NATSAL-4 (2022-2024) still pending publication

Unfortunately, there are no relevant publications on sexual behavior for the latest survey NATSAL-4. Although publications are expected over the next few years.

Our first paper (Hawes et al., 2010) doesn't have the latest NATSALs but covers papers reviewing age at first sex. Given its age it's now more a historical review of changing virginity post sexual revolution.

In recent decades, the age at which young people become sexually active has fallen (Schubotz et al., 2004; Wellings & Field, 1996; Wellings et al., 2001). Median age at first intercourse was 17 years among the 40- to 44-year-old age group in NATSAL 2000 for both men and women (Wellings et al., 2001). The proportion of women reporting first intercourse before 16 years has increased over recent decades, although

not since the mid 1990s (Wellings et al., 2001). Studies also report on the gender differential in timing of first sexual intercourse, and several studies (Blenkinsop et al., 2004; Currie et al., 2004; Currie & Todd, 1993; Lenciauskiene & Zaborskis, 2008; Wight et al., 2008), but not all (Wellings et al., 2001), have found a higher proportion of girls having had sex by age 16, compared with boys.

Our second paper (Mercer et al., 2013) looks at NATSALs 1 through 3. They provide nice tables with overviews. We will go through the relevant points so you don't have to. All findings are age adjusted.

• Lifetime sexual partners increased substantially for women, but less so for men. These changes plateaued between 2000 and 2010.

• Number of partners in the last year changed between 1990 and 2000 for men and women but changes ceased 2000 to 2010.

• Sexual frequency substantially declined across all surveys.

There are two more tables that cover cohort effects by sex within the NATSAL-3.

• A larger proportion of younger cohorts report losing their virginity before 16 than older cohorts. This is substantial and significant for both men and especially women.

• The number of permanent virgins has remained mostly constant at ~2% for men and ~0.5% for women.

• Owing to the sexual revolution, as implied for the previous table, partner count has generally increased for females.

The last paper we'll look at (Mitchell et al., 2023) for NATSAL includes the most recent data from COVID.

Partner count dropped between 2010 and COVID, but this could be a COVID effect.

Compared with 10 years previously (Natsal-3), women and men in Natsal-COVID-2 were less likely to report two or more sexual partners in the past year (women: 5.4% vs 13.5%, aOR 0.30, 95% CI 0.25 to 0.36; men: 9.6% vs 19.4%, aOR 0.37, 95% CI 0.31 to 0.43). Similar differences between surveys were observed for numbers of reported new sexual partners for women and men, and for WSW (lesbians) and MSM (gays).

However, as was with all successive NATSALs, sexual frequency continued to decline.

Verdict: Incels falling/stagnation.

• Sexual frequency: Declining consistently and significantly between surveys.

• Partner counts: Increasing over the decades, especially for women.

• Age of first sex: Generally declining across age groups.

Australia

Australia (SSASH)

The Australian National Survey of Secondary Students and Sexual Health or SSASH, established in 1992, has had seven iterations up until 2021.

The latest survey was conducted online, recruiting respondents through social media advertising:

For the 2021 survey, data were collected ... via an online instrument that took an average of 33 minutes to complete.

Promoted via social media advertising ... participants were recruited from Instagram (47.0%, n = 2,740), while 39.0% (n = 2,272) were from Facebook and 7.0% (n = 409) from TikTok ... The final sample included the 6,841 ... 4,459 participants who completed the entire survey.

Given this, it's critical to ensure the sample is representative of the general population. However, it appears they had disproportionately high lesbian, gay, bisexual and female samples:

Wilson et al. (2020) estimate the percentage of the Australian population who identify as LGB to be approximately 3.5%.

The percentage of young people identifying as LGB in this survey is high ... 23.3% (n = 1,586) identified as bisexual, 6.0% (n = 406) as gay or lesbian.

The largest group to respond to the survey were ... female (n = 4,456, 65.1%) ... (27.8%) who identified as male.

While this survey's methodology has limitations in representativeness, its large sample size and longitudinal nature still provide valuable insights. My worry is that variance between survey samples might be quite high to obfuscate trends

The results contradict the US narrative. The Australian youth is enjoying more sex than ever while their incel cousins cope and seethe.

Australia (ASHR)

The Australian Study of Health and Relationships or ASHR, is an Australian wide decennial survey conducted in 2001-2002 (ASHR1), 2012-2013 (ASHR2) and 2022-2023 (ASHR3).

According to their website they claim that analysis and publication was planned for 2024. Now that 2024 is over, they have clearly failed to meet this timeline.

However, there were plenty of papers covering cohort differences between ASHR1 and ASHR2.

The survey is more representative and better sampled than its sister online survey. Quoting from (Rissel, Heywood, et al., 2014)'s brief description of the method.

... computer-assisted telephone interviews were completed by a representative sample of 20,094 Australian residents aged 16–69 years from all states and territories ... Respondents were selected using dual-frame modified random-digit dialling (RDD), combining directory-assisted, landline-based RDD with RDD of mobile telephones. The overall participation rate among eligible people was 66.2%.

8577 completed the long- form interview, and 11 517 completed the short-form interview.

Reading through four papers covering the differences between ASHR1 and ASHR2, (Badcock et al., 2014), (Rissel, Heywood, et al., 2014), (Rissel, Badcock, et al., 2014), and (de Visser et al., 2014). Here were the best quotes and summaries regarding differences across birth cohorts and between surveys.

From the paper 'Characteristics of heterosexual regular relationships among a representative sample of adults', between ASHR1 and ASHR2, there was a decline in sexual frequency.

Finally, respondents in ASHR2 reported a significantly lower average frequency of sex in the past 4 weeks than respondents in ASHR1 (both sexes, P < 0.001). For men, the average frequency of sex declined from 1.86 (CI: 1.76–1.96) to 1.50 (CI: 1.41–1.59) and in women, it declined from 1.82 (CI: 1.72–1.92) to 1.52 (CI: 1.42–1.63).

In 'First vaginal intercourse and oral sex among a representative sample of Australian adults', as one can see from Table 6, the proportions of 16-19 year olds reporting ever having sex did not significantly change for men (p=0.07) and for women (p=0.28).

The paper provides cohort data on the median age of first sex (virginity). It seems like there's a notable post-sexual revolution decline in virginity then a plateau.

The results show a significant decline in the median age of first vaginal intercourse for both men and women among those born between the 1940s and the 1960s, but no further decline since.

For all cohorts, the median age at first sex increased for men, and female lifetime partner count increased. Plausibly the observed trends here stem from the oldest pre-sexual revolution cohorts dying, with post-revolution generations replacing them between surveys.

Verdict: Incels falling/stagnation.

• Sexual frequency: Significant declines in ASHR.

• Partner counts: General increase in women. Otherwise mostly unchanged.

• Age of first sex: Declined in older cohorts.

France

France (CSF)

The introduction to Contexte des Sexualités en France, translated by Google, provides a good overview of France's large sex surveys over time.

A first scientific survey on the sexual behavior of the French was conducted in 1970 among 2,600 people by Pierre Simon and his team (Simon 1972). It was followed in 1992 by the survey “Analysis of sexual behavior in France”, conducted among 20,000 people, under the responsibility of Alfred Spira and Nathalie Bajos (Inserm), then by the survey “Context of sexuality in France” conducted in 2006 among 12,000 people under the responsibility of Nathalie Bajos and Michel Bozon (Ined) in 2006. The last two surveys, initiated and funded by the National Agency for Research on AIDS (ANRS), have given rise to numerous publications (Spira and Bajos, 1993, Bajos et al.; 1998, Bajos and Bozon 2008). The results have helped guide the development of sexual health policies, particularly in the area of ​​HIV infection. The new survey “Context of sexuality in France” (CSF-2023) required 5 years of work. Initiated in the fall of 2019, its schedule was delayed by the Covid-19 pandemic. This document is the first presentation of the methodology and the first results. They are presented for mainland France.

TL;DR, France has four main surveys conducted in 1970, 1992, 2006 and 2023 respectively.

There's a paper published on the first three surveys (Bajos et al., 2010), and the document summarizing the latest survey (Bajos et al., 2024).

In the latter paper, which focuses on the latter three surveys (1992 - 2023, not 1970), they plot the median age virginity is lost by cohort. As we can observe, and as we have seen in studies from other countries, the decline in age of first sex is greatest for women.

Notably, in the recent decade, age of first sex has started to rise again! This is important because the CSF 2023 has data for the last decade unlike the Aussie and British surveys.

Both papers provide two tables on total lifetime partner counts. However, the first paper includes the 1970s survey and excludes 2023s survey, while the second document does the opposite. For convenience, the numbers from both were gathered in one accessible plot.

As can be observed, for women there has been a universal increase in lifetime sexual partner count. Male lifetime counts were mostly stagnant 1970-2006 with the exception of the 2023 survey. The 2023 survey looks suspicious. Keep in mind these are averages not medians.

Finally, and most interestingly, we have sexual frequency broken down by single and partnered status!

The survey finds something that I didn't expect, sex frequency is declining not just overall, but within couples and singles! As I was writing this article, I expected that most of the variance in the decline of sexual frequency could be attributed to the decline in marriage, however, it appears that this is not the case.

The decline it seems is less pronounced over a four week period, than over a year.

Verdict: Incels paradox. Less sex than ever before but with more partners.

• Sexual frequency: As measured by sexual activity in the past year, a universal decline across all surveys, even adjusting for relationship status. When only looking at mean sexual encounters in the past 4 weeks, declines only appear in the latest survey.

• Partner counts: Increasing over the decades, especially for women. This holds true even for the oldest surveys.

• Age of first sex: Generally declining across age groups since the sexual revolution, with the exception for the last decade.

Other Countries

Our primary task was to investigate if US sexlessness trends were comparable to culturally or developmentally similar countries. For this, France, UK and Australia are sufficient, but not exhaustive. The rest of the countries here are low hanging fruit; but have not been systematically reviewed.

Japan

Japan has the 'National Fertility Survey', which has been conducted roughly every 5 years since the 1980s. Its structure is different from the other countries with different aims, as a result, it separately questions never-married and married persons.

There's only one figure in the summary report pertaining to sexlessness broadly. It largely confirms what we observed with France, UK and Australia in that sexual experience seems to have increased for women since the sexual revolution. Other than that, there haven't been many significant changes.

One could argue that declining marriage could be causing a selection effect in the older age groups, however, this cannot be a large effect for the teenagers sampled, as few are married by that age.

Italy

A study on Italian University students looked at the correlates of virginity in surveys from 2000 and 2017. They find virginity rates falling, which is not surprising given the results we've seen particularly in the UK.

Scandinavia

A Scandinavian study (Hansen et al. 2019) measured sexual behaviors among 100k women in Sweden, Denmark and Norway.

Mostly the same observations as other countries, that lifetime partner counts had increased among women.

Our findings show a largely stable debut age among the cohorts born during the 1960s and 1970s, with median age at first intercourse of 16 for most of the Danish cohorts, and 17 for most of the Norwegian and Swedish cohorts... Our study further demonstrated a shift towards earlier first intercourse in younger female birth cohorts (born 1983-1994), to a median debut age of 16 in all countries

The overall median (interquartile range) number of lifetime sexual partners among women in Denmark, Norway, and Sweden were 6 (3-10), 5 (2-10), and 6 (3-11), respectively... the proportion of women who reported >10 sexual partners was higher in 2012 than in 2005 in all age groups and in all countries

The Rest

The NATSAL had a 'scoping review' where in the appendix it conveniently lists population sex surveys from around the world. Countries not covered that are listed here include New Zealand, Ireland, Slovenia, South Africa, and Spain. There's also a study on Germany finding sexual declines. From memory Nuance Pill had a post with a similar summary, however, I've not been able to locate it.

It should be noted that Durex, the condom company, conducts well resourced international surveys of sexual behavior across dozens of countries every year or so. I could only find three reports that they published online; and it would seem this would be an excellent resource, but it seems mostly closed source.

Conclusion, Discussion

Is Declining Sexual Frequency Mostly Declining Marriage?

An interesting phenomena or perhaps even a paradox is that the number of lifetime sexual partners is increasing, the age of intercourse is falling, there's more sexual liberty and acceptance of sexual minorities than ever, however, despite all this, sexual frequency is declining.

My simple theory is that most of the decline in sexual frequency can be reduced to the decline of relationships and marriage.

Very simply, using the GSS we can measure the standardized age and sex adjusted difference in quantified sexual frequency. The difference is small, but still significant at -0.267 standard deviations.

As was seen in the French data, there's also a slightly significant effect by birth cohort. In the GSS it's about -0.00204 standard deviations per year, or across 50 years (e.g. born in 1950 vs 2000) about -0.102 standard deviations.

So ballpark estimate a total difference of about 0.3 standard deviations (ignoring the multiple assumptions we just made) could be explained between the oldest surveys with the oldest cohorts and youngest cohorts in the latest surveys. Again, this is due to a small cohort effect and a marriage effect.

This theory makes sense because declining marriage rates wouldn't affect the age at which one loses their virginity. Furthermore, less serious long term relationships substituted for short term flings would result in higher lifetime partner counts too - which is exactly what we see happening in the data. However, once again, the effect of declining marriage seems insufficient to account for the reported declines.

Closing thoughts

This review has been enlightening. Judging just from the US data which shows multi decadal declines in sexual behavior across multiple datasets, I was anticipating finding roughly the same trends. However, the conclusions were more nuanced. Men and especially women are losing their virginities sooner, have more sexual partners and greater diversity of sexual experiences, but less often.

The data reveals several interesting paradoxes. While sexual liberation has continued its march forward - with decreasing ages of first sexual experience and increasing lifetime partner counts - actual sexual frequency has declined consistently across surveyed nations. This pattern holds true even when controlling for relationship status, suggesting deeper societal shifts beyond just declining marriage rates.

The French data is particularly revealing, showing sexual frequency declining not just among singles but within established couples too. This contradicts the simple narrative that modern dating apps and changing relationship dynamics are solely responsible for observed changes in sexual behavior.

Notably, the trends differ markedly between genders. Women have seen the most dramatic changes since the sexual revolution, with substantial increases in lifetime partner counts and earlier sexual debuts. Male patterns, by contrast, have remained relatively stable outside of the overall frequency decline.

Annoyingly, the UK and Australian datasets haven't published their findings despite sitting on it for over a year. What's worse is that directly on their website they promised publication last year. These newer datasets could be crucial for understanding post-2020 trends, especially given the French data suggesting possible accelerated declines in recent years. With increasing digitization and changes in social patterns, we might be seeing the emergence of new behavioral patterns that older datasets can't capture.

The international perspective ultimately challenges many common explanations for changing sexual behavior. Popular theories about dating apps, smartphones, or gender dynamics struggle to explain why different countries show such varying patterns despite similar technological and social changes. This suggests we need more nuanced frameworks for understanding these trends rather than universal explanations based solely on US data.

• A 2005 paper by Huebner claimed worldwide innovation was declining—particularly after 1850. It was frequently cited to support theories about IQ dysgenics, but appears to be flawed.

• By analyzing the same historical text Huebner used but counting notable figures instead of innovations, no significant decline was found.

• When comparing against a comprehensive database of historical figures, the data actually shows an increase in innovation-related figures over time, contradicting Huebner's findings.

• The apparent decline in Huebner's study likely stems from selection bias in his source material (a single history book) rather than reflecting actual innovation trends.

Introduction

During the late 2010s and early 2020s, IQ dysgenics theory gained considerable attention within certain academic and online communities. Michael Woodley had steadily published research on estimating IQ dysgenics using Galton's methods, while Edward Dutton promoted these ideas through his YouTube channel and co-authored work "At Our Wits' End: Why We're Becoming Less Intelligent and What It Means for the Future". The prevailing narrative suggested that individuals with lower cognitive abilities reproduced at higher rates and earlier ages than their more intelligent counterparts. Given the established heritability of IQ, proponents argued that humanity faced inevitable intellectual decline—with the satirical film Idiocracy serving as an unintended prophecy.

However, subsequent research has substantially revised this narrative. New evidence published over recent years, combined with the reduced prominence of key advocates in academic discourse, has led to a more nuanced understanding of these phenomena. Recent comprehensive analysis by HBD researchers has revealed two critical findings:

First, IQ dysgenics effects are considerably smaller than initially proposed. When IQ is measured directly in dysgenics calculations—rather than using proxies like educational attainment—the effects, while statistically significant, are modest in magnitude.

Second, selection pressures against educational attainment exceed those against IQ itself. Since educational attainment frequently served as a proxy for IQ in genetic and biobank studies, previous estimates of cognitive decline were substantially inflated.

During the peak of dysgenics discourse, this supposed cognitive decline provided explanatory frameworks for various societal trends. Innovation represented a particularly compelling case study: proponents argued that declining innovation rates directly reflected population-level IQ dysgenics. The cornerstone evidence for this argument came from Huebner's 'A Possible Declining Trend for Worldwide Innovation', which produced the influential and frequently cited (152 times) figure:

This finding aligned seamlessly with dysgenics theory, which posited that cognitive decline began when natural selection pressures against lower IQ relaxed—presumably during the post-industrial period. The year 1850 represented an ideal inflection point for this narrative.

The apparent strength of this correlation prompted Michael Woodley to publish additional research exploring this relationship (cited 62 times). He reported a remarkably strong correlation of r = .876 across 56 decades of data:

However, this correlation appears problematic when viewed through the lens of current understanding and may represent circular reasoning rather than genuine causal relationships. The logical circularity operates as follows: IQ dysgenics theory predicts that cognitive decline should manifest as innovation decline; Huebner's paper appears to document innovation decline; this decline is then cited as evidence supporting IQ dysgenics theory. Yet if Huebner's innovation decline stems from methodological artifacts rather than genuine historical trends—as this analysis suggests—then using his findings to validate dysgenics theory constitutes circular reasoning. The correlation becomes spurious: researchers identified a dataset that appeared to confirm their theoretical predictions, but the dataset itself may reflect selection bias rather than the phenomenon it purports to measure.

Nevertheless, the central question remains: could Huebner's 2005 paper genuinely document innovation decline independent of IQ dysgenics theory? This analysis addresses this possibility by examining whether the sources underlying Huebner's conclusions suffer from systematic selection bias.

Methodology: Replicating and Extending Huebner's Approach

Huebner derived his innovation data from 'The History of Science and Technology' by Bryan Bunch, a chronologically organized reference work spanning from 3000 BC to 2003 AD. The book employs a systematic format where years appear in page headers or footers, with multiple innovation events on single pages distinguished by clearly marked temporal divisions.

Rather than replicating Huebner's innovation counting approach, this analysis focuses on the historical figures mentioned in the same source material. This methodological shift enables examination of whether apparent innovation decline reflects genuine historical trends or artifacts of the source material's scope and selection criteria.

Data Processing Framework

The analysis employed Natural Language Processing (NLP) techniques to systematically extract and analyze historical figures from Bunch's text:

1. Extract complete text from all pages

2. Apply regular expressions to identify year references for each page

3. Use NLP algorithms to distinguish personal names from common nouns

4. Construct a comprehensive database linking page numbers, text content, temporal references, and identified names

To transform raw name extraction into a robust historical dataset, several refinement steps were necessary:

1. Calculate temporal associations by averaging the years associated with each distinct name across the entire book

2. Integrate findings with the cross-verified database of notable people, 3500BC-2018AD, which encompasses all figures documented in Wikipedia and Wikidata—providing a far more extensive historical record. The database was filtered to exclude contemporary celebrities and athletes

3. Apply Levenshtein distance algorithms to match names from Bunch's text with entries in the cross-verified database, optimizing matches based on temporal proximity between book references and documented lifespans

This process identified approximately 6,000 historical figures mentioned in Bunch's work between 1450 and 2003. While the matching algorithm achieved high accuracy for most entries, some matches involved figures with variant spellings or slight temporal discrepancies. The overall correlation between extracted names and verified historical figures was sufficiently robust for analytical purposes.

The cross-verified database integration served two essential functions: it provided occupational classifications for figures mentioned in Huebner's source material, and it supplied precise birth and death dates for temporal analysis. This enhanced dataset enabled systematic examination of several critical hypotheses.

Research Questions

This analysis addresses four key questions regarding potential biases in Huebner's source material:

1. Does 'The History of Science and Technology' exhibit declining mentions of historical figures per capita, paralleling Huebner's innovation decline findings?

If per capita mentions of innovators declined alongside innovations, this would support Huebner's conclusions, since innovation requires innovators.

2. When examining only occupations represented in Huebner's source, do these same occupations show decline in the comprehensive cross-verified database?

This test reveals whether 'The History of Science and Technology' systematically samples from declining occupational categories while neglecting emerging fields.

3. Does the cross-verified database demonstrate decline among the most eminent figures (top 0.1%), and do such patterns correlate with trends in Huebner's source?

This addresses whether observed patterns result from the source material's focus on exceptionally prominent figures, or whether the comprehensive database oversamples recent, less historically significant individuals.

4. Do occupations overrepresented in 'The History of Science and Technology' correlate with declining trends in the broader historical record?

Positive correlation would indicate systematic bias toward occupations experiencing historical decline.

Results

Historical Figures Per Capita in Huebner's Source

Analysis reveals no significant decline in living historical figures per capita within 'The History of Science and Technology' comparable to Huebner's reported innovation decline. The data shows relatively stable or slightly increasing trends rather than the dramatic post-1850 decline that characterized Huebner's innovation metric.

Occupation-Specific Trends in Comprehensive Database

Examination of the comprehensive cross-verified database yields results directly contradicting expectations of decline. When filtering for occupations represented in Huebner's source material, the data reveals substantial increases rather than decreases over time. This finding suggests that apparent plateaus in 'The History of Science and Technology' result from incomplete sampling rather than genuine historical trends.

Elite Figure Analysis

Analysis of the top 0.1% most eminent figures in the cross-verified database closely mirrors the broader occupational trends, effectively eliminating concerns that observed increases result from oversampling less significant recent figures. The consistency between elite and general population trends strengthens confidence in the underlying data quality.

Furthermore, when filtering the elite figures specifically for those in Discovery/Science fields—the categories most relevant to innovation—the same increasing pattern emerges, directly contradicting Huebner's decline narrative for the most accomplished scientists and discoverers.

The eminence of these top-tier figures is readily apparent from their universal recognition:

Occupation Bias Analysis

Systematic examination of the 100 most frequently mentioned occupations in 'The History of Science and Technology' reveals no correlation between their representation in Bunch's work and their historical trajectory in the comprehensive database. This finding rules out the possibility that the source material inadvertently focuses on occupations experiencing genuine historical decline.

Discussion and Conclusions

This comprehensive analysis demonstrates the absence of declining trends in eminent historical figures across both Huebner's source material and the larger cross-verified database. The stability of figure mentions in Bunch's work, contrasted with clear increases in the comprehensive database, cannot be attributed to occupational bias or selective focus on elite individuals.

These findings suggest that Huebner's reported innovation decline likely reflects the inherent limitations of his source material. Bunch's 'The History of Science and Technology,' while representing a substantial scholarly effort at 785 pages, contains a sample size approximately 100 times smaller than the cross-verified database. Comprehensive representation of historical innovations proportional to the actual number of innovators would require a reference work of several thousand pages—an impractical undertaking for any editorial team.

The nature of innovation itself has undergone fundamental transformation since Huebner's analysis. Traditional occupational categories like 'inventor'—frequently sampled in Bunch's work—have indeed declined, but this reflects structural changes in how innovation occurs rather than absolute innovation decline. Contemporary innovation manifests through incremental improvements following predictable trajectories, exemplified by exponential cost reductions in technologies like batteries that follow Moore's Law-like patterns. Modern innovation operates through accumulated marginal gains rather than the dramatic singular breakthroughs that characterized earlier eras and dominated historical narratives.

Additionally, current understanding of dysgenics effects as modest rather than catastrophic undermines theoretical foundations for expecting per capita innovation decline. Without substantial population-level cognitive decline, arguments for innovation decline based on human capital deterioration lack empirical support.

The convergence of methodological limitations in Huebner's source material, evolving patterns of innovation, and revised understanding of cognitive trends provides substantial evidence for reconsidering his conclusions. Two decades of subsequent research have developed far more sophisticated approaches to measuring innovation across multiple dimensions, rendering single-source historical analyses increasingly obsolete.

This analysis supports a verdict of substantial doubt regarding Huebner's innovation decline thesis, suggesting that his findings reflect sampling artifacts rather than genuine historical trends.

Bunch, B., & Hellemans, A. (2004). The history of science and technology: A browser's guide to the great discoveries, inventions, and the people who made them from the dawn of time to today. Houghton Mifflin Harcourt.

Dutton, E., & Woodley of Menie, M. A. (2018). At our wits' end: Why we're becoming less intelligent and what it means for the future. Imprint Academic.

Huebner, J. (2005). A possible declining trend for worldwide innovation. Technological Forecasting and Social Change, 72. https://doi.org/10.1016/j.techfore.2005.01.003

Jensen, S. (2024, January 30). Are we getting dumber? Data on IQ and fertility across the world. Center for the Study of Partisanship and Ideology. https://www.cspicenter.com/p/are-we-getting-dumber

Laouenan, M., Bhargava, P., Eyméoud, J. B., Gergaud, O., Plique, G., & Wasmer, E. (2022). A cross-verified database of notable people, 3500BC-2018AD. Scientific Data, 9(1), 290. https://doi.org/10.1038/s41597-022-01369-4

Woodley of Menie, M. (2012). The social and scientific temporal correlates of genotypic intelligence and the Flynn effect. Intelligence, 40, 189-204. https://doi.org/10.1016/j.intell.2011.12.002

TL;DR

• Analysis of three major US datasets (YRBSS, NSFG, GSS) confirms increasing sexlessness and later loss of virginity among youth.

• Relative sexual inequality is increasing (Gini coefficient) due to an increase in virgins, but absolute inequality is declining.

• Reasons for being a virgin appear to be increasingly involuntary or ambiguous, rather than voluntary.

Introduction

Over the past few years, discussion of "the rise of sexlessness" has reached fever pitch in certain corners of the internet. You've probably seen this infamous plot making the rounds:

Despite being outdated and analytically shallow, this chart continues generating clicks and heated debates. While serious researchers have long moved past such sensationalism, there's still a gap in the literature that needs addressing.

As far as I can tell, no one has comprehensively examined all three major US datasets on this topic simultaneously. Analysts typically focus on either the NSFG or GSS, with the YRBSS rarely getting attention. This fragmented approach leaves room for cherry-picking and incomplete conclusions.

What we really need is comprehensive international meta-analysis covering not just sexual behavior, but related developmental milestones: first employment, substance use, social relationships. Think Jean Twenge's work on iGen, but broader in scope and geographic coverage. Since that ambitious project exceeds my current resources, I'm focusing specifically on sexlessness trends within the United States.

Note, I did attempt an international analysis a few months later anyway

This geographic limitation comes with important caveats. American trends don't automatically generalize globally, and the evidence I've reviewed suggests rising sexlessness may be uniquely American rather than a universal phenomenon among developed nations.

It's tempting to assume that social changes in the US eventually spread to other Anglosphere countries and beyond, and this pattern does hold for many trends like declining fertility, certain cultural movements, etc. But we should resist reflexive assumptions about universality without proper cross-national evidence.

Overview

This analysis draws on three complementary US datasets, each offering distinct strengths and limitations:

1. Youth Risk Behavior Surveillance System (YRBSS)

2. National Survey of Family Growth (NSFG)

3. General Social Survey (GSS)

These figures reflect sample sizes after filtering for relevant variables. Sample size alone doesn't tell the whole story; the age ranges and time periods each survey covers matter enormously for statistical power. When we account for both dimensions, the effective density per age-year cell works out to roughly 1,218 for YRBSS, 118 for NSFG, and just 15 for GSS.

This makes the GSS appear weakest statistically, trailing the YRBSS by nearly two orders of magnitude. But the GSS offers unique advantages that justify its inclusion: it's the only survey measuring sexual frequency (beyond just partner counts) and provides crucial data on older generations. These capabilities prove essential for understanding broader demographic patterns.

Each dataset illuminates different aspects of American sexual behavior. The YRBSS excels at tracking youth trends over time, the NSFG captures detailed sexual histories during prime reproductive years, and the GSS offers the broadest age range with the longest historical perspective.

Sexlessness in the YRBSS

Let's start with the YRBSS, which provides our longest-running view of youth sexual behavior. The central question: is sexlessness actually rising?

The answer appears to be a resounding yes. What's particularly striking is that the trend is accelerating faster among males than females, a difference that reaches statistical significance, as we'll see in the regression results below.

To get a clearer picture of timing, I used a spline model to predict virginity rates by age, then calculated the exact age at which virginity drops below 50% of the sample. Think of this as the median age of virginity loss:

The trend is remarkably consistent: approximately six months later every decade. While this might seem modest, it represents a meaningful shift in the timing of sexual debut across generations.

Inequality in the YRBSS

Beyond simple trends in sexlessness, we can examine how sexual experiences are distributed across the population. This requires distinguishing between two types of inequality: absolute and relative.

Absolute inequality measures raw differences using metrics like variance or mean absolute error. Relative inequality, on the other hand, uses functions like the Gini coefficient that economists favor because they capture proportional differences.

Here's why this distinction matters: Imagine 10% of the population owned all the wealth in a society. If that 10% doubled their wealth, the Gini coefficient wouldn't budge, but the absolute gap between rich and poor would double. Conversely, if the wealthy lost half their money, the Gini would stay constant even as absolute differences shrank. Keep this in mind as we examine the data.

Now, let's apply this framework to sexual behavior. What patterns do you see here?

This looks suspiciously similar to our virginity plot. Is that a coincidence? Let's check the correlation:

The correlation is nearly perfect at r = 0.996. This tells us something crucial: virtually all the rising relative inequality stems from increasing virginity rates, not from changes in behavior among sexually active students.

To dig deeper, let's examine lifetime partner counts broken down by decile, grade, and sex, computing each group's deviation from the overall mean. Values closer to zero indicate that group's average is closer to the population mean.

Here's the paradox: despite massive increases in relative inequality, absolute inequality is actually declining across nearly all groups.

There's one notable exception: men in the 10th decile among 12th graders show increasing deviation. This reflects a data artifact: the highest response category caps at "6 or more partners." When overall sexual activity declines, this ceiling effect makes the top decile appear to diverge more from the mean, even though their actual behavior may not be changing much.

The Lorenz curve provides another illuminating perspective on these trends:

The inequality shift is stark. Among 12th grade males in 2023, the top 10% account for 50% of all partner counts, double the 25% figure from 1991. Perhaps more telling: back in 1991, even students in the bottom 30% were having some sexual experiences. Today, roughly half report none at all.

YRBSS Conclusions

This analysis suggests we can actually steelman the Incel argument. The grievances may indeed stem from rising inequality, but it's specifically relative inequality that matters, not absolute trends. Consider the perspective of today's average high school graduate: he's now likely to be a virgin, watching a smaller group of peers account for most sexual activity. This concentration simply didn't exist a generation ago.

The YRBSS data reveals several clear patterns:

1. Rising sexlessness among youth

2. Increasing relative inequality in sexual experiences

3. Declining absolute inequality (the paradox mentioned above)

One common Incel claim, that women are completely unaffected by these trends, proves false. However, the data does show women are less affected than men in both sexlessness and virginity trends.

The statistical models confirm these observations:

Sexlessness: ``r lm(virgin ~ year * sex_coded + year * grade_coded, data = virgin) %>% broom::tidy() %>% arrange(p.value) ``

The fitted coefficients:

Mean Virginity: ``r lm(Virgin Age ~ year * Sex, data = threshold_ages) %>% broom::tidy() ``

The fitted coefficients:

While these YRBSS results aren't overwhelmingly significant, they point consistently in one direction. The crucial question becomes: do these patterns replicate across other datasets?

Sexlessness in the NSFG

When I originally examined these trends, NSFG data only covered through 2019, and the sexlessness patterns appeared relatively modest. The 2023 data release changed that picture dramatically.

For years, the NSFG's substantial sample size advantage over the GSS made it a favorite reference point for those dismissing claims about rising sexlessness. That changed with the latest release, which shows unmistakable increases across multiple metrics.

The old arguments against these trends have lost their foundation:

First, while the NSFG covers the shortest time period of our three surveys, making trend detection more challenging, the recent patterns are too pronounced to ignore. Second, the YRBSS actually maintains a larger overall sample size. Most importantly, the newest NSFG data clearly demonstrates rising sexlessness regardless of these methodological considerations.

The visualizations tell a consistent story:

Inequality in the NSFG

The inequality patterns in the NSFG mirror those we observed in the YRBSS:

This represents a clear replication of the youth trends in an older, broader population.

Some might argue for caution in interpreting these recent NSFG declines, given that 2022-23 represents the first post-COVID data collection. I disagree with this hesitation for several reasons.

The YRBSS and GSS have been documenting these same trends for decades. The broader pattern was already evident before this latest NSFG release. These new results simply eliminated the final counterargument that could be mounted against the underlying trends.

That said, the NSFG's user guide does advise caution when comparing 2022-23 data to earlier years. The survey shifted from face-to-face interviews to a multimode design, and COVID-19 prevented proper experimental evaluation of this methodological change. These factors suggest potential artifacts may be influencing the apparent trends. Nuance pill wrote an article examining these methodological concerns.

While virginity increases prove statistically significant when examining all age groups combined, the trends within individual age cohorts (20, 25, 30) don't reach significance independently.

Notably, the widening gender gap in virginity rates that we observed in the YRBSS doesn't appear in the NSFG data.

``r nsfg_stnd %>% group_by(age, year, sex) %>% mutate(age = round(age/5, 0) * 5) %>% summarise(virgin = mean(not_virgin == 2, na.rm = TRUE)) %>% filter(age <= 30) %>% mutate(age = as.factor(age)) %>% lm(data = ., virgin ~ year * age + year * sex) %>% broom::tidy() ``

The fitted coefficients:

This again illustrates the absolute/relative inequality distinction: the youngest cohorts show the highest relative inequality yet the lowest absolute deviations.

Reasons for Remaining Virgin

The NSFG provides unique insight into motivations for virginity through several variables: stated reasons for remaining virgin, how much respondents wanted their first sexual experience when it occurred, and sexual orientation patterns.

The most striking pattern here is religion's decline as a reason for virginity. This directly contradicts the Institute for Family Studies' argument that "we're having less sex because of religion" (Stone, 2021). Their analysis suggested that increasing religiosity among youth was driving sexlessness trends.

But this gets the causation backwards. If sexual activity is declining universally, we'd naturally expect religious people to have less sex too, not because religious attitudes are strengthening, but because the overall trend affects everyone. The data clearly shows religious motivations for virginity are actually weakening over time.

The potentially involuntary reasons, 'Other' and "Haven't found the right person", are both increasing substantially.

If these responses truly reflect involuntary virginity, we should see them become more common with age.

The age patterns support our interpretation. 'Other' increases dramatically with age, exactly what we'd expect if it represents involuntary circumstances. "Haven't found the right person" remains relatively constant across the twenties; if this reflected a clear voluntary choice, we might expect it to vary more systematically with age (possibly decline). Religious reasons decline in older age groups, likely reflecting both marriage patterns among religious individuals and broader generational shifts in religiosity.

The classification challenge remains that most stated reasons appear voluntary rather than involuntary. Only "Haven't found the right person" clearly suggests involuntary circumstances. The 'Other' category is inherently ambiguous, but given that most major voluntary reasons are explicitly covered elsewhere, this residual category likely captures predominantly involuntary situations.

Desire to Lose Virginity

We can test whether 'Other' truly represents involuntary virginity by examining the WANTEDSX variable, which measures how much respondents wanted their first sexual experience when it occurred. If men are becoming increasingly desperate to lose their virginity as they age, we'd expect to see heightened desire among those who finally do have sex.

The patterns here are revealing, though the rape statistics for women represent a deeply troubling aspect of these data. Setting that aside, men show relatively constant levels of desire across ages. To quantify this more precisely, I converted responses to a continuous scale: 'I really wanted' = 1, middle response = 0.5, and "I didn't want" = 0.

Men and women appear to converge in their reported desire for first sexual experiences. If anything, men maintain consistently high levels of wanting across all ages, while women show some increase with age.

This pattern somewhat contradicts the hypothesis that 'Other' represents an increasingly desperate demographic, at least for men. The data suggests men maintain constant eagerness regardless of age, rather than showing the escalating desperation we might expect from prolonged involuntary celibacy. For women, there may be more support for this interpretation.

The Asexuality Question

Could rising asexuality explain the increase in 'Other' responses? The NSFG captures sexual attraction through the ATTRACT variable:

I classify responses 6-9 as potential indicators of asexuality or sexual uncertainty.

While these responses do increase with age as expected, asexuality can't account for the majority of 'Other' virginity reasons. At most, these categories cover about 15% of virgins, while 60% select 'Other' as their reason. This suggests most 'Other' responses reflect circumstances more consistent with involuntary celibacy rather than asexuality.

Breaking this down by year and sex reveals potentially interesting patterns, though we must interpret cautiously given extremely small sample sizes (median n ≈ 15 per data point):

There does appear to be an upward trend worth monitoring as more data becomes available.

Incels Rising

Having examined the nuances of different virginity reasons, I can now classify them along the voluntary/involuntary spectrum:

While many of these "involuntary" classifications remain ambiguous, they're more likely involuntary than not given the exclusion of major voluntary reasons.

The trends by year tell a striking story:

Involuntary reasons for virginity appear to be increasingly dominating among men, with women following a similar but less pronounced pattern. This aligns with certain aspects of Incel theory, which often characterizes female virginity as more voluntary than male virginity.

Age patterns reinforce this interpretation:

As expected, involuntary reasons become more prevalent with age across both sexes.

NSFG Conclusions

While increases in sexlessness and partner counts don't reach statistical significance in the NSFG, they consistently point in the expected direction across multiple measures.

More importantly, the data clearly indicates a shift from voluntary to involuntary reasons for virginity, even accounting for the ambiguous nature of some classifications.

The declining role of religion offers one potential explanation for rising involuntary celibacy. For previous generations, religious commitment provided a socially acceptable framework for remaining virgin until marriage, a choice individuals could embrace with genuine conviction. As religiosity declines among younger Americans, this cultural safety net has largely disappeared, leaving sex as an unmarked norm without traditional taboos.

This cultural shift may leave sexually unsuccessful men particularly vulnerable. Without religious justification for abstinence, the natural male inclination toward sexual pursuit becomes more prominent but remains unsatisfied, potentially contributing to feelings of involuntary deprivation.

This remains speculative, but the correlation between declining religious explanations and rising involuntary reasons deserves serious consideration.

Sexlessness in the GSS

For readers who've made it this far, the General Social Survey needs little introduction. Its unique strengths include an extensive age range, diverse variables beyond sexuality, decades of historical data, and crucially, measures of sexual frequency rather than just partner counts.

Starting with sexual frequency patterns:

The trends are unmistakable: lower frequency responses are becoming more common while higher frequencies decline. This represents a clear shift toward reduced sexual activity across the population.

Annual partner counts show an even starker pattern:

The 'No Partners' category shows the strongest trend of any response. Note that bottom-row data points should be disregarded due to extremely small sample sizes (approximately n=10 each).

For measuring virginity, the GSS offers one viable variable: total sexual partners since age 18. The proportion reporting zero partners provides our best proxy:

This represents the most dramatic increase of any measure we've examined. Following the linear regression trend, we're looking at a doubling over thirty years, definitively refuting claims that "the GSS doesn't show declining sexlessness."

Inequality in the GSS

To examine inequality patterns, I converted the categorical SEXFREQ and PARTNERS variables into continuous measures. Sexual frequency was quantified as instances per week using these conversions:

The relative inequality trends match our expectations:

Sexual frequency shows clear increases in relative inequality over time.

For absolute inequality measures:

Particularly for sexlessness, the first quintile appears to be converging toward the mean (zero), suggesting declining absolute inequality even as relative inequality rises.

Is Everyone Affected?

The GSS's broad age range allows us to examine whether these trends affect all demographics or concentrate in specific age groups:

While relative inequality appears to be rising across age groups, the individual variable trends remain somewhat ambiguous, making it difficult to definitively conclude whether all demographics are equally affected.

Where are the Incels?

Given rising sexlessness across all three datasets, we might expect parallel increases in Incel-related online activity. To test this, I examined US Google Trends data using Elliot Rodger searches as a proxy for Inceldom:

Surprisingly, the trends remain relatively flat. This disconnect could reflect two possibilities: either Incel culture's brief history makes it too recent to capture longer-term sexlessness patterns, or we're observing a selection effect.

The selection effect explanation seems more plausible. Early Incels represent a highly specific demographic: disproportionately mentally ill, NEET (Not in Education, Employment or Training), and likely possessing higher-than-average libidos.

As sexlessness spreads to broader populations, however, the newly sexless individuals differ markedly from this prototype. Through simple regression towards the mean, they're more likely to be employed, mentally healthier, and less obsessively focused on sex.

This demographic shift means that most newly sexless individuals probably won't be drawn into the same terminally online ecosystems that captured earlier Incels. This represents genuinely good news; it suggests that rising sexlessness may not translate proportionally into the growth of these communities.

In a more sophisticated analysis, the population of various manosphere communities was measured (Horta Ribeiro et al., 2021).

Between 2016-2019, Incel communities did show growth, and the broader manosphere appears to have expanded since its inception. However, several factors complicate this interpretation:

• Internet user populations have grown massively, so we need to consider whether activity has increased relative to overall internet usage

• The data reflects international patterns rather than specifically US trends

• Selection effects may mean that those most predisposed to Incel ideology were simply early internet adopters

Conclusions

The evidence from US data reveals several consistent patterns:

• Youth virginity rates are increasing while sexual frequency and partner counts decline

• The NSFG initially appeared to contradict these trends, but the 2023 release brought it into alignment with both YRBSS and GSS findings

• Sexual milestones continue to be reached, but at progressively older ages

• Stated reasons for virginity are shifting from voluntary to involuntary

• Relative sexual inequality is rising dramatically

• Absolute inequality, paradoxically, continues to decline

These findings suggest that American sexual behavior is undergoing a significant transformation, with implications that extend far beyond simple statistics about sexual activity.