IA rereading: submitted to 200 articles already published in Nature, Science or Cell, four LLM recalculate three quarters

Researchers have blindly read articles published in 2026 in major journals: models demand massive revisions, contradict each other and often miss errors that have been deliberately slipped.

The lack of rereaders is driving more and more voices to entrust part of peer review to language models. A team led by Bastien Le Guellec wanted to see how these models judge work that has already crossed the most demanding editorial filter. It collected 200 articles published in 2026 under the CC BY 4.0 licence in twenty general science, biology and clinical medicine journals, including Nature, Science and Cell, and then asked four recent LLMs (two owners, two open-weight) to write a rereading report for each, without revealing their source.

Severe, inconsistent and unreliable

Of 800 reports, 605 (75.6%) recommend major revision or rejection. The differences between models are considerable: according to the model, this proportion ranges from 17.5% to 100%, on articles already accepted. The authors then introduced into the manuscripts a methodological or statistical flaw specific to each discipline: according to the model, it is only detected in 17% to 60.5% of cases.

The models therefore combine two defects: they require corrections on validated works and allow errors placed before their eyes to pass. Researchers conclude that their opinions must be read as an uneven signal, very dependent on the model and the formulation of the request, and not as a verdict on scientific validity.