An estimated 6.5% to 16.9% of conference peer reviews were substantially changed by AI — does that make them sound alike?

This explores whether LLM help with writing peer reviews and papers makes scientific writing sound more alike. The corpus doesn't contain a direct measurement of vocabulary or style narrowing in reviews, and it's better to say so than to fake one. What it does have answers two smaller questions: how much review text is touched by LLMs, and what forces push writing toward a single style.

Start with how much text is affected. One analysis of reviews from ICLR, NeurIPS, CoRL and EMNLP estimates that between 6.5% and 16.9% of review text was substantially modified by an LLM How much peer review text shows signs of LLM modification?. The method is worth noticing. It doesn't flag individual reviews. It measures a shift in word patterns across the whole pool of reviews. So the field's best detection tool already treats LLM writing as something that changes the overall shape of the text, not as a feature of single documents. The same study found the most LLM modification among reviewers who were less confident, rushed, or less engaged. Those are the reviews where a reviewer's own voice was least likely to show up anyway.

Why would that text converge? The corpus points to the generation process itself. Token prediction pulls output toward the center of the training data. It doesn't push toward competing positions, so the result is smooth prose that adds claims without adding new perspectives Does LLM generation explore competing claims while producing text?. A related note shows that the same model writes in recognizable registers: a sycophantic chat voice and a 'falsely objective' published-prose voice, each inherited from its training data Why do LLMs produce such different writing in chat versus posts?. A peer review is exactly the kind of text that calls up the second voice: measured, evaluative, impersonal.

The less obvious finding is that readers and machines both reward that convergence, so it reinforces itself. Expert readers couldn't reliably tell LLM-written abstracts from human ones. LLM-edited abstracts got the highest clarity ratings and were preferred 55% of the time even when authorship was disclosed Can readers tell LLM abstracts from human ones?. On the machine side, rewriting a paper's rhetoric while keeping its science unchanged measurably shifts LLM reviewer scores How much does rhetorical style shift AI review scores?. LLM judges also respond to surface cues like formatting and authority signals Can LLM judges be fooled by fake credentials and formatting?. Put these together and authors have a reason to write in the style that machine reviewers and human skimmers both rate well. That narrows style through incentives, not only through copying.

There is a check on the simplest version of this story. The apparent preference of LLM-assisted reviewers for LLM-written papers disappears once paper quality is controlled for Do LLM reviewers actually favor LLM-written papers?. A randomized ICML trial found that banning LLM use, versus allowing limited use, barely changed scores or decisions Does banning LLM use in peer review change review outcomes?. So the evidence points to style converging without clearly changing outcomes. That makes it a quieter problem than bias: reviews may come to sound alike while deciding roughly the same things. To answer whether they actually sound alike, someone would need to measure vocabulary diversity directly, and this corpus doesn't yet contain that work.