AI audited the AI literature — 99.2% of papers flagged

For as long as there has been modern science, acceptance at a peer-reviewed venue has been the certificate of trust: experts read it, it passed, so it is true. That certificate just received its first large-scale independent audit — and the auditors were AI agents. This week, three separate efforts put machines to work checking the published record of machine learning itself: a company that tried to rerun every oral paper at ICML 2026, a community challenge that put a third of the conference under an agent microscope, and a GPT-5-powered checker that flagged 99.2% of 2,500 top-venue papers for at least one objective error. Science has a new gatekeeper, and it reads faster than any human ever will.

The machines that came for peer review

The most comprehensive picture comes from a study by researchers including NEC Labs: "To Err Is Human: Systematic Quantification of Errors in Published AI Papers via LLM Analysis." They built a Paper Correctness Checker on GPT-5 and ran it over 2,500 papers from ICLR (2018–2025), NeurIPS (2021–2025) and TMLR (2022–2025), counting only objective mistakes — errors in formulas, derivations, calculations, figures and tables with verifiable ground truth, not novelty or writing quality. The checker found an average of 4.66 mistakes per paper; 99.2% were flagged for at least one. Math errors dominated at 54%, and the trend is worsening: the average NeurIPS paper went from 3.8 flagged mistakes in 2021 to 5.9 in 2025, a 55.3% rise. Human experts confirmed 263 of 316 machine flags as genuine — 83.2% precision. Most flagged issues are minor, but 30.8% of NeurIPS papers and 23.8% of ICLR papers carried at least one substantive error that could change how results are interpreted.

Then there is the audit that reran the experiments. In an analysis posted on July 22, SAI Labs — a for-profit research-review company in Delaware — deployed AI agents against all 168 oral-presentation papers at ICML 2026, as Nature reported this week. The agents extracted central claims, downloaded code and data, reran experiments where possible and compared outputs with what the authors reported. Of the 92 papers with at least five testable claims, only 34 had more than 40% of those claims reproduced by the agents — and only 8 cleared 80%. James Evans, the University of Chicago sociologist who worked on the audit, put the economics in perspective: reproducing every experiment in a top ML paper cost a median of $8,900. "Peer review has never been a sufficiently strong verifier to facilitate cumulative assembly," he concluded.

Some findings look worse than sloppiness. One paper boasted of training only 0.77% of a base model's parameters, but its released checkpoint trained 6.31% — an eightfold difference. Another showed a reliability table from an evaluation model absent from its open-source code. Four ICML papers depended on models since taken offline, unreproducible by anyone, ever. None of this requires fraud: failures ranged from missing key files and broken dependencies to outputs that don't match the paper.

Why now — and why the literature can't hide

The audits land now because the verification arithmetic finally works. Peer reviewers never had time to download data, run models and recheck every conclusion, and the volume has outgrown them: ICLR submissions grew from roughly 1,013 in 2018 to 19,619 in 2026, and a modern paper involves hundreds of fiddly details of code, data and settings. Agents changed the cost curve. Nature's report — headlined "AI agents are checking the scientific literature — and spotting decades-old errors" — shows the effect reaches far beyond machine learning. Sebastian Pios, a theoretical chemist at Zhejiang Lab in Hangzhou, was using AI to predict molecular boiling points when the model produced values clashing with a 75-year-old reference database. He assumed the AI was wrong; checking the original literature showed the database was wrong. The model went on to catch a typo in an old paper and incorrect century-old boiling-point measurements that had propagated into the scientific canon.

The ICML effort's scale ends any claim that verification is boutique. Hugging Face and AlphaXiv ran the Agent Reproduction Challenge from July 15 to August 2, in which more than 1,200 participants pointed coding agents at accepted papers to reproduce or refute claims with real executed experiments, logged publicly and scored by an automated judge built on GLM-5.2. By the end, more than 2,000 papers — a third of the conference — had been reproduced or falsified across roughly 13,000 repositories and 3 terabytes of artifacts.

The fix: papers built for machines

The emerging answer is a different artifact, not better proofreading. A group of 37 researchers from Stanford, Michigan, Carnegie Mellon and MIT, led by Jiachen Liu, argues in a paper covered by IEEE Spectrum that the PDF itself is the problem. They identify a "narrative tax": the tangled exploration tree of a real project gets pruned into a clean success story, dead ends and failed hypotheses vanishing. And an "engineering tax": the details that determine whether experiments run — model versions, environments, hyperparameters, preprocessing — are scattered across appendices and repositories, or never recorded. Their analysis of PaperBench's 8,921 expert reproduction requirements found only 45.4% fully documented in paper PDFs. Their proposed replacement, the Agent-Native Research Artifact (ARA), packages research as a machine-operable bundle with four layers: scientific logic, executable code, an exploration map preserving failed routes, and evidence linking every claim to raw results.

The numbers favor the new format, at least in ML. Across 450 understanding questions, agents answered correctly 93.7% of the time from ARA packages versus 72.4% from PDFs plus code; on failed-approach questions, the gap was 81.4% versus 15.7% — PDFs simply don't contain that information. On 150 reproduction tasks across 15 papers, ARA won 64.4% to 57.4%, the advantage growing on harder tasks. Across 24,008 agent runs, 90.2% of compute went into failed exploration — dead ends ARA preserves so the next agent skips them. One irony the authors surely appreciate: the ARA paper itself is still a PDF.

What the skeptics say — and what to watch

Hold the panic, partly: 83.2% precision means roughly one in six flags is a false positive, and most confirmed errors are minor typos, not collapsed results. "Unreproducible" is a spectrum from fraud to unpinned dependencies, and conflating the two inflates the crisis. Auditing AI with AI inherits hallucination risk: in the ARA team's own 15-paper test, the traditional group fabricated results twice, the ARA group once — machines still invent output, and a checker can confidently "correct" a correct paper. And the audits so far cover code-tractable ML venues; wet-lab science is a far harder problem.

The direction of travel, though, is clear. Post-publication machine verification is becoming a standard layer of science, reshaping incentives: authors who know agents will rerun their experiments will document more carefully, top-venue acceptance loses some halo as a quality certificate, and an erratum economy is forming — Chinese coverage gleefully frames the audits as a free source of research topics. Watch whether ICML and NeurIPS require machine-checkable artifacts at submission, whether SAI Labs-style audits become a paid standard with real editorial power, and whether error rates plateau once authors optimize for the checker — the classic Goodhart trap. The machines have read the literature. Now the literature has to decide whether to be written for them.

If AI becomes the real peer reviewer, does a top-venue acceptance still mean anything? Tell us in the comments.

Sources: Nature News · To Err Is Human (arXiv) · NEC Labs · James Evans on LinkedIn · ICML 2026 Agent Repro Challenge (Hugging Face) · Hugging Face results livestream · IEEE Spectrum · Agent-Native Research Artifacts (arXiv) · QbitAI · 36Kr English