← Back to feed

A study published in Nature Human Behaviour on August 24, 2026 puts specific numbers on the linguistic homogenization effect of large language models. The findings have direct implications for hiring assessments, clinical psychology, and any domain that infers human characteristics from written text.

The paper — Sourati, Karimi-Malekabadi, Ozcan, et al. — analyzed more than 880,000 texts across Reddit stories, news articles, ArXiv papers, essays, social media posts, and political speeches. The researchers combined observational analysis of real-world writing trends with controlled experiments in which GPT-3.5, Llama 3 70B, and Gemini Pro rewrote thousands of human-authored texts.

What the Numbers Say

LLM rewriting reduced variation in writing complexity by 21 to 50 percent depending on the dataset and model. The reduction was consistent across all seven datasets and held regardless of the rewriting instruction used: whether the prompt asked for clarity, fluency, polish, academic tone, or readability, the homogenization effect appeared every time.

Semantic content was largely preserved. In 87 percent of cases, the original and rewritten texts had meaning-similarity scores above 0.95. The LLMs flattened style while keeping substance intact.

The downstream effect: classifiers trained to infer personal characteristics from writing were six percentage points less accurate on LLM-rewritten texts than on originals. The signal was there in the original writing. After an LLM pass, it was not.

Which Identity Signals Disappear

The study identified specific lexical associations that LLM rewriting weakens:

• Pronoun use linked to extraversion
• Friend-related words linked to loyalty
• Future-focused words linked to age

Some associations survived rewriting:

• Negative-emotion words linked to neuroticism
• Religion-related words linked to moral purity
• Social words linked to gender

The pattern is not random. LLM-rewritten texts consistently shift toward writing styles associated with older, male, politically liberal authors — with positive moral valence and lower expressed empathy. The researchers attribute this bias to the RLHF training process. Reinforcement learning from human feedback reduces harmful content by preferring certain stylistic registers, and those preferences appear to carry over into rewrites even when the original content is neutral.

The Observational Evidence

Beyond controlled experiments, the paper tracked real-world writing trends across ArXiv, Reddit, and a news dataset. All three showed a measurable inflection point after ChatGPT’s public launch in November 2022: variance in linguistic complexity features (Vocabulary Simpson Index, Type-Token Ratio, Hapax Legomena) declined, while mean complexity increased. The interpretation is that LLMs are pulling writing toward a narrower, more polished center.

Granger causality tests confirmed the effect persists over time — it is not a one-time shock from the November 2022 release, but a sustained compression of the diversity distribution.

Practical Consequences

The research identifies four areas where the homogenization creates measurable risk:

Clinical psychology and mental health. Language patterns are used to detect early indicators of depression, anxiety, and cognitive decline. If patients increasingly write through LLMs, those signals degrade.

Hiring and recruitment. Writing samples used to assess candidates’ communication style, cultural fit, and personality become less informative after LLM editing — and the remaining signal is biased toward the demographic characteristics the training process favored.

Personalization. Recommendation systems and personalized services that infer preferences from language will see reduced accuracy as user-generated content is polished through the same small set of models.

Cultural preservation. Writing style carries community and generational markers. A persistent homogenization effect slowly erodes those distinctions at the corpus level.

Why the Timing Matters

The study used GPT-3.5, Llama 3 70B, and Gemini Pro for its experimental conditions. Those models are now the lower tier. Claude Fable 5.1, GPT-5.6, and Grok 4.6 are the current frontier, and they are embedded in Google Docs, Microsoft Word, Slack, and email clients used by hundreds of millions of people daily. The scale of the homogenization effect the paper measured is almost certainly larger now than the datasets from which it was estimated.