WHAT AI ERASES WHEN IT REWRITES

Sept. 2026

How much can we trust AI-generated knowledge of the world—and the concepts we use to understand it—when it draws on texts already reshaped by other AI systems?

A paper published on August 24 in Nature Human Behaviour brings together three studies and over 880,000 texts (Sourati et al. 2026).

In the rewriting experiments, models largely preserve the content but reduce stylistic diversity. Reported statistically significant reductions in the variance of linguistic complexity range from 21% to 50%. This does not, of course, mean an equivalent decline in intelligence.

The more revealing result comes next. After rewriting, systems that infer authors’ characteristics from their texts become less reliable. Some judgments shift systematically: authors are more often classified as male or older, for example. The person has not changed; the cues used to form a picture of them have.

Rewriting can also change what a text claims. Analyzing 4,900 scientific summaries produced by ten models, Peters and Chin-Yee (2025) found that most models tested generalized findings beyond their original scope. Drop a qualification, and a finding that holds only under certain conditions becomes a general claim.

Consider a hypothetical example. Employees choose their start time but cannot negotiate their ten-hour working day. A text criticizes this limited autonomy. Summarized as “employees choose their working hours,” it loses the very point it was making. A critique of a constraint becomes a description of freedom.

If another AI uses this summary as a source, it might present the company as an example of employee autonomy. The original objection would be missing. Meaning lost in the writing could then distort what the system claims to know.

The hypothesis goes beyond profiling: systems could explain the world using distinctions other systems retained, never encountering those they erased. The concepts would remain—freedom, justice, suffering—but some of the thinking they made possible would be lost.

Knowledge could become more coherent by reproducing the same simplifications, without becoming any more faithful to the world.

Liviu Poenaru

References

Peters, Uwe, and Benjamin Chin-Yee. 2025. “Generalization Bias in Large Language Model Summarization of Scientific Research.” Royal Society Open Science 12 (4): 241776. doi:10.1098/rsos.241776.

Sourati, Zhivar, et al. 2026. “The Shrinking Landscape of Linguistic Diversity in the Age of Large Language Models.” Nature Human Behaviour, August 24. doi:10.1038/s41562-026-02550-0.