A study published in Nature Human Behaviour examines a question that has been floating at the edges of the AI writing-tool discussion without much empirical weight behind it: what actually happens to the statistical properties of text when LLMs help humans write it? The answer, across 880,000 texts and seven datasets, is that variance collapses — and it collapses reliably.
The core measurement is writing-complexity variance. The researchers operationalize this through a suite of features — sentence length distributions, syntactic depth, lexical diversity, readability scores — and track how that variance changes when texts are revised with LLM assistance. The reduction is 21–50% depending on dataset and model, and it holds at statistical significance (p ≤ .05) across all conditions. That’s not a marginal effect.
More interesting than the raw compression is what gets erased. Variance in writing complexity is not random noise; it carries signal. The patterns associated with a writer’s gender, estimated age, political ideology, and moral values — established categories in forensic linguistics and computational stylometry — are suppressed in LLM-polished text. The models amplify characteristics associated with dominant patterns in their training data while flattening deviation. The result is text that resembles the center of whatever distribution the model has been trained to reward.
This is exactly what you’d expect if you think about how these models were trained. RLHF-optimized models are steered toward text that human raters evaluate as high quality. High quality, in practice, means fluent, clear, appropriately formal for context, free of awkward constructions. Those are genuine virtues. But “fluent and appropriately formal” is itself a style — one associated with educated, majority-language, professionally-acculturated writers. When a model nudges text toward that style, it nudges away from the styles that deviate from it: colloquial, regional, idiosyncratic, syntactically unusual for deliberate effect.
The researchers test seven datasets spanning social media posts, news articles, and scientific writing. The homogenization pattern holds across all three domains, which suggests it’s not an artifact of a particular genre’s conventions. They also test across multiple LLM variants and multiple prompt formulations, and the effect is consistent. This is not a model-specific bug; it’s a property of the class of models.
The practical implications span several fields that probably haven’t fully internalized this yet. Forensic linguistics uses stylometric features to identify individual writers, attribute authorship, and estimate demographic properties — techniques that matter in legal proceedings and fraud detection. If the writing process increasingly routes through a common editing layer, those tools lose resolution. Hiring decisions that use writing samples as a signal are measuring something different than they were. Clinical psychology uses written text as a window into cognitive and emotional state; that window is being tinted.
There’s also an implication that runs back into the AI development loop. Web text scraped as training data is increasingly LLM-assisted — this is not speculative, it’s visible in output statistics from multiple tracking sources. Training on that corpus means training on compressed signal: text where the idiosyncrasies that make individual voices distinguishable have been smoothed toward the statistical center. The question of how that compounds across training generations has not been definitively answered, but several theoretical treatments suggest the variance reduction accumulates.
The paper paired with it in the same Nature Human Behaviour issue addresses a slightly different framing — the blurring of personal identity through AI writing assistance specifically — and reaches compatible conclusions. Together they constitute something closer to a definitive empirical statement than the prior literature of individual studies on specific models or tasks.
None of this argues that LLM writing assistance shouldn’t exist. It argues that it should be understood as a system that, by design, moves text toward a particular stylistic center — and that the costs of that movement fall unevenly across writers whose styles deviate most from that center.