In 1984, George Orwell imagined a language engineered until "no one speaks any other." More than 800 million people now use LLM writing assistants like ChatGPT and Claude to draft, polish and rewrite their words. A study published this week in Nature Human Behaviour suggests Orwell's direction was right — just slower and stranger than he predicted. It isn't that AI bans words. It's that it quietly makes everyone sound like the same person.
Led by researchers at the University of Southern California (USC), the study analyzed more than 880,000 texts — news articles, academic papers, argumentative essays, social media posts and speeches — and ran three experiments probing how LLM-assisted writing changes language. The headline: AI writing assistants are measurably flattening linguistic diversity, and the flattening is systematic, not random.
Writing complexity is converging
The first study compared corpora before and after ChatGPT's launch in November 2022. Across arXiv preprints, Reddit posts and local news, the variance in writing complexity narrowed significantly. Style spreads didn't shrink because the world got more uniform; they shrank in step with the adoption curve of LLM "polishing."
The second study made the mechanism explicit. In controlled rewriting experiments, LLM rewrites reduced the variance of writing complexity by 21% to 50%. Distinct drafts by different people, run through the same assistant, came back more alike than they went in.
The "default persona" bias
The third study is the one that should worry psychologists. The team paired text corpora with psychometric data — age, gender, Big Five personality, empathy, moral values — and trained classifiers to predict those traits from language. Then it compared predictions on original text versus LLM-rewritten text.
- Prediction accuracy dropped by an average of roughly 6 percentage points (F1). Age was hit hardest: F1 fell from 0.351 to 0.260, a 26% loss.
- The change is not random noise. Rewritten text is systematically judged as older, more moralizing, less empathic, less extraverted — and more open and agreeable.
- The bias held across different LLMs and different rewriting prompts. Whatever the model, rewriting pushes everyone toward the same "default persona."
In other words, LLMs don't erase identity signals so much as overwrite them with one profile. The classic word-to-person associations that linguists and social scientists rely on are being selectively eroded: some associations vanish, others survive.
Why language is becoming the "final form"
The mechanism isn't mysterious. LLMs are trained to produce the statistically most likely continuation, and the most likely text is, by definition, the most mainstream text. When hundreds of millions of people route their writing through that filter, individual voice — the specific rhythm, word choice and stance that makes a sentence yours — is the first thing to be averaged away.
Worse, the loop feeds itself. AI-assisted text circulates back into training corpora; the next generation of models trains partly on its own homogenized output, accelerating convergence. What looks like "AI getting better at writing" is partly a closed loop: the models, the users and the corpus are all pulling toward the mean together.Public open-training runs like Marin 535B make this pipeline visible: the taste of the training data becomes the taste of the model. A fuller breakdown of the three experiments — corpus sizes, causality tests and the eroding word-person links — is in our earlier analysis.
What's actually at stake
This isn't purely an aesthetic loss. Language-based signals are infrastructure for real systems: psychologists screen for depression from textual cues, employers audit hiring language for fairness, marketers personalize at scale, social scientists read values from speech. If those signals flatten toward one default profile, three things happen:
- Tools degrade quietly. Mental-health screening and personality inference tuned on pre-LLM text lose accuracy as real-world text becomes more alike.
- Fairness gets a new blind spot. If everyone is rewritten toward the same persona, classifiers stop seeing difference — which sounds "fair" until you remember they were also the tools that detected bias.
- Culture thins at the edges. Register, dialect and community idiom don't survive "most likely continuation" averaging. What's disappearing isn't just word count; it's the texture of how communities reason.
Print expanded registers; telegraph compressed them; digital writing fragmented them into niches. The LLM assistant is the first technology in history that actively homogenizes prose at population scale — and it does so one polite suggestion at a time.
What to do about it
- Writers: use AI as an editor, not a ghostwriter. Keep a human draft before any model pass; the voice you preserve is data for everything downstream.
- Prompt for variety: ask for specific registers, sentence rhythms and non-default vocabulary. The bias is a default, not a law.
- Product teams: style-preserving modes, explicit provenance labels on AI-assisted text (verifiable watermarking is becoming standard), and sampling strategies that resist the mean are becoming a user-respect feature, not a niche one.
- Researchers: track corpus contamination the way you track data leakage. The validity of every linguistic-signal study published in the next decade depends on it.
As one of the USC researchers put it: "Languages, traditions, ways of thinking — once they disappear, we can't get them back. The question isn't whether this is difficult to fix. It's whether we can afford not to."
Source: USC Dornsife; paper published in Nature Human Behaviour.
