Report 034 · AI in the Lab
At ICLR 2026, 21% of the reviews were written by a machine. A Stanford-led team then showed you can raise those scores by having an LLM rewrite your paper, mostly by changing how it sounds. Peer review just acquired an attack surface.
I have been on both ends of peer review. I have submitted work in microwave spectroscopy and waited on reviewers, and I have been the reviewer, reading someone else's methods section at eleven at night trying to decide whether their calibration holds up. It is a slow, uneven, frequently annoying process, and it is also the only quality filter science actually has. So it is worth knowing that a meaningful share of it is now being done by language models, and that a team just demonstrated how easily those models can be steered.
The number comes from a study by Joachim Baumann, Jiaxin Pei, Sanmi Koyejo, and Dirk Hovy, presented in July 2026 at the International Conference on Machine Learning. They looked at "all 75,800 reviews from the 19,490 papers under review at ICLR 2026," one of the largest venues in machine learning, and found that "15,899 reviews (21%) are AI-generated."
Not AI-assisted. Generated. Roughly one review in five at a flagship conference was produced by a model rather than by the human being who agreed to evaluate the work. And this is not a fringe behavior: a survey of 1,600 scientists across 111 countries, reported by Science News, found that more than half had used AI tools to help review papers. The reviewers are drowning in submissions, and they reached for the obvious tool. I understand the impulse. That does not make the output a measurement.
The interesting part is not the count. It is what the team did next, which they call paper laundering. They took 60 randomly selected ICLR 2026 papers spanning a wide range of research areas, had an LLM rewrite each one, and then fed the original and the rewritten version to AI reviewers to see whether the score moved. They ran 24 conditions in total: four zero-shot prompts, two rewriting models, three reviewer models (GPT-5.1, GPT-5.4, and Claude Sonnet 4.5).
It moved. Their finding, verbatim: "AI review scores are trivially gameable through paper laundering: prompting an LLM to rewrite a paper could significantly increase the scores from AI reviewers." The overall mean increase was +0.45 points on a 1-10 scale, statistically significant at p<0.001 in nearly every condition. Half a point does not sound dramatic until you remember that acceptance decisions at these venues routinely turn on a tenth.
And here is the part that should bother anyone who cares about science rather than about publishing: the gains did not come from better research. The team reports that "laundering disproportionately makes stylistic modifications, with increased hedging words" and emphasis words. Science News, covering the work, reported that the rewritten papers also included obvious cases of scientific misconduct, with models fabricating experimental findings, and the paper itself notes some edits were more substantive, involving "hallucinated AI slop." So the reviewer's score went up when the prose got hedgier and more confident-sounding, and it did not reliably go down when the model invented results.
This is where I default to thinking about it as a measurement problem, because that is my training. An AI reviewer is an instrument. You point it at a paper and it returns a number that is supposed to stand in for scientific quality. Before you trust any instrument's reading, you ask two things: does it respond to the quantity you care about, and does it respond to things you do not care about?
On the second question the answer is now documented: it responds strongly to style. On the first, the study has a direct measurement. Testing how well scores predicted actual acceptance, human scores hit an AUC of 0.822 while AI scores managed only 0.710. The machine is a noticeably worse predictor of the outcome it is being used to substitute for. That is a calibration failure, and running more papers through it does not fix a calibration failure, it just produces more confident wrong numbers.
There is a second, subtler failure. The team found what they call a "hivemind effect of excessive agreement within and across papers that reduces perspective diversity." Concretely, GPT-5.1's reviews of original papers showed 37.4% higher cross-paper similarity than human reviews, and within-paper similarity rose 8.7%. That matters because the entire design logic of peer review is that you get three or four different people, with different blind spots, and disagreement is the signal. Three reviewers who converge because they share a base model are not three reviewers. They are one reviewer, sampled three times, and the appearance of consensus is an artifact of the sampling.
My other career was spent on the other side of this exact problem. In electronic warfare you spend a great deal of time on automated systems that classify a signal and make a decision about it, and you learn a rule the hard way: any automated classifier that scores on surface features rather than on the underlying thing can be spoofed by someone who knows which features it scores on. That is not a bug in a particular system. It is the general property. It is why spoofing a GPS receiver works so well: the receiver checks whether a signal has the right shape, not whether it is telling the truth, so a well-formed lie sails through with full confidence.
An LLM reviewer that rewards hedging language is a receiver checking shape. And the moment a scoring system becomes consequential, it starts getting optimized against. Right now this is a research demonstration by people trying to warn the field. It will not stay that way, because the incentive to get a paper accepted is enormous and the cost of running your draft through a rewriting prompt is about four cents.
Worth being precise, because this is easy to flatten into "AI bad." The team is not calling for a ban. Their conclusion is that "addressing the peer review crisis requires a science of peer review automation, not general-purpose LLMs deployed without rigorous evaluation." The crisis is real: submission volume at these venues has outrun the supply of willing expert reviewers, and something has to give. Their argument is that if you are going to automate an evaluation, you have to validate the automated evaluator against the thing it replaces, the way you would qualify any new instrument before putting it in the workflow. Nobody swaps a spectrometer into a production line because it produces numbers faster. You run it against knowns first.
The recurring lesson on this beat is that a model is very good at reproducing the surface of a thing and much worse at the thing itself. It reproduced protein structures that mostly did not fold, and weather forecasts that flatten the records. Here it reproduces the surface of a review: the right length, the right hedges, the right tone of expert reservation, without the part where a human being who has actually run that experiment notices the thing that is wrong.
So the practical question, whether you are a researcher, a lab director, or anyone deploying AI as a judge of anything: has this evaluator been validated against the outcome it stands in for, and do you know what it responds to besides that outcome? For AI peer review, as of July 2026, the published answers are 0.710 versus a human 0.822, and it responds to your writing style. Know that before you trust the number, in either direction. If a machine reviewed your paper and liked it, that is weaker news than it feels like. And if it disliked it, that is weaker news too.
U.S. Army combat veteran (Counter-IED / Electronic Warfare), peer-reviewed researcher in microwave spectroscopy, and founder & CEO of Shroombiosis. Consults on laboratory operations, AI, and supplement formulation.