In their 2024 ICML paper, Weixin Liang, Zachary Izzo, and coauthors developed a method to estimate AI-modified text in large collections, applying it to peer reviews from four major AI conferences post-ChatGPT launch. They found 6.5% to 16.9% of review text showed signs of substantial AI modification. This highlights risks in consequential settings like peer review, where AI-generated content could distort evaluation.

A related 2024 Nature study by Ilia Shumailov and coauthors showed that training models on recursively generated data leads to model collapse, losing rare data patterns and narrowing output. The author tested three low-cost checks—perplexity, near-duplicate similarity, and embedding density—on a movie review dataset from Mendeley (2019) to detect AI-generated text. At a setting catching 80% of AI-generated reviews, embedding density flagged 47% of human-written source reviews.

Filtering using the combined score of all three checks reduced sentiment classifier accuracy from 67.5% to 57.5% on a held-out set. The human-written label was inferred from source (IMDb) and pre-ChatGPT publication date, as the collection does not verify individual authorship. These results suggest current text-based checks may over-flag human content and harm downstream model performance when used naively for filtering.