LLMs Erase Linguistic Diversity and Identity Signals in Human Writing

New paper in Nature Human Behaviour shows LLMs reduce how linguistically diverse people are when they write with them — and strip out the identity signals embedded in our words.

Paper: arXiv:2502.11266 — The Impact of LLM Use on Linguistic Diversity and Identity Signals in Human Writing (Nature Human Behaviour, 2025)

What it finds

Two things:

1. LLMs reduce linguistic diversity. When people write with LLMs, their writing becomes less varied — less rich in the range of words, phrases, and constructions they use. The language flattens toward what the model tends to produce.

2. LLMs erase identity signals. The words we use carry information about who we are — our background, our community, our habits, our personality. LLMs, by smoothing toward a generic "good" style, strip out those signals. The writing becomes more "correct" or more "average," and in doing so loses the fingerprints of the person behind it.

In short: LLMs don't just change what we say — they change who we sound like when we say it. The paper's punchline: LLMs strip out the soul of our writing.

Why it matters

This isn't just about style. If the words we use are signals of identity, then widespread LLM-assisted writing could quietly erode the diversity of voices we hear — not through censorship, but through optimization toward a generic center. The more we write with these tools, the more the distinctiveness of individual and community voices gets sanded off.

It's also a useful corrective to the "LLMs just help you communicate better" framing — they help you communicate more conventionally, which is not the same thing.

Related

• ai-graduates-employment — how AI is reshaping professional writing and communication
• agent-psychosis — the psychological side of human-AI interaction
• anthropic-distillation-attacks — where the model's "generic" style comes from and why it matters

By

Natalie Shapira et al. (Nature Human Behaviour, 2025)