AI Agents Are Beginning to Audit Scientific Knowledge

While using AI to predict molecular boiling points, Zhejiang Lab theoretical chemist Sebastian Pios found results that conflicted with a 75-year-old reference database. He initially suspected the model, but checking the primary literature showed that the database was wrong in this case. Nature reported two similar findings: a typo in an older paper and incorrect century-old measurements that later entered reference works.^[S1]

The reliable step was not to trust the AI automatically. The disagreement became an anomaly signal, a researcher returned to the primary source, and only then was the error identified. AI did not overturn established science; it made a contradiction visible enough to investigate.^[S1]

At the same time, verification is expanding from individual values to full papers. SAI agents extract central claims, retrieve code and data, rerun experiments where possible and compare outputs with reported results. Google's Paper Assistant Tool and Stanford's Paper Correctness Checker place verifiers at different points: before submission and after publication.^[S2][S4][S3]

From one anomalous value to batch replication

SAI analyzed 168 ICML 2026 oral papers and completed full replication attempts for 105. Its current public report states that 92 of those papers contained at least five judgeable claims; 34 reproduced more than 40% of attempted claims, while eight exceeded 80%.^[S2][S1]

The 8-of-92 figure is the easiest number to misuse. It does not mean only eight papers were correct. Automated replication depends on claim extraction, artifact retrieval, environment reconstruction, code execution and result matching. Missing files, broken dependencies, unavailable data or checkpoints, and agent execution errors can all lower the score.^[S2][S1]

Pipeline failure is still informative. SAI reports common problems including code that does not run as released, underspecified experiments and outputs that differ from the paper; four papers depended on models that had already been retired. The audit primarily measures how fragile the path from manuscript narrative to verifiable result remains.^[S2]

Start with errors that have relatively clear answers

Federico Bianchi, James Zou and colleagues deliberately narrowed the Paper Correctness Checker to objective errors in formulas, derivations, calculations, figures and tables. It does not judge novelty, importance or writing quality. From a random sample of 60 papers with at least one flagged substantive mistake, the paper's own authors reviewed all 316 flagged issues by hand and confirmed 263, for 83.2% precision. Recall, measured separately on 90 planted mistakes, was 60.0%; recall on real mistakes was not measured.^[S3]

In that analysis, the mean detected errors per NeurIPS paper rose from 3.8 in 2021 to 5.9 in 2025, a 55.3% increase. This does not prove that AI made papers worse; paper volume, complexity, research practices and reviewer workload also changed. It shows that consistent automated checking makes longitudinal audits possible. The paper itself treats the counts as a conservative lower estimate: with recall around 60%, missed mistakes are not counted.^[S3]

Verifiers are moving upstream, before publication

Google offered the Paper Assistant Tool to STOC and ICML authors as private pre-submission feedback, not as a replacement for formal review. Among 733 ICML survey respondents, 35.4% said it identified substantive theoretical gaps requiring meaningful work, while 31% ran new experiments because of the feedback.^[S4]

Verifier systems therefore occupy two positions. Before publication, they help prevent new errors from entering the literature; after publication, they audit papers and databases already inside the scientific corpus. One resembles quality control, while the other pays down accumulated verification debt.

AI4S has focused heavily on generation: new molecules, proteins, materials and hypotheses. Scientific knowledge, however, is not the sum of generated candidates; it is a set of claims that can be checked, reproduced and corrected. If generation accelerates while verification throughput remains fixed, verification debt will accumulate.^[S4]

Pios's case also shows that literature and databases are scientific infrastructure. A value moves from an experiment into a paper, handbook and database, then into later experiments and models. The longer it propagates, the further it drifts from its original provenance. Making old knowledge auditable supports not only reviewers but the foundations of training data and scientific decisions.^[S1]

An AI verifier does not need to replace the best expert to be useful. If it can scan a large corpus at low marginal cost, reliably surface a subset of anomalies and hand supporting evidence to researchers, it already changes the economics of scientific verification.^[S1]

Nature combines cases at different maturity levels.

The boiling-point case is a researcher account reported by Nature without a corresponding primary publication. SAI is a company-published online audit, while the Paper Correctness Checker, PAT and SPOT are preprints. They do not form one comparable measure of AI-verifier accuracy.^{[S1][S2][S3][S4][S5]}

The five sources side by side: study, object, sample, metric, check

Nature's boiling-point case.

Object: one value in a reference database. Sample: one case. Metric: none. Check: a researcher returning to the primary literature. It is an illustrative case (an AI disagreement led a person to an error) and measures no system's accuracy.^[S1]

SAI's audit.

Object: all 168 ICML 2026 orals; 105 full replication attempts, 92 with at least five judgeable claims. Metric: share of attempted claims reproduced per paper, reduced-scope runs included. Check: comparison with the paper's reported results, and its review comments against OpenReview reviews (it covered 78% of points that two or more human reviewers agreed on). The replication results were not checked by people, and SAI says the system is in beta and can make mistakes.^[S2]

Paper Correctness Checker.

Object: objective mistakes in 2,100 ICLR and NeurIPS papers plus 400 TMLR papers. Metrics: 83.2% precision (60 papers, 316 flags, checked by the authors) and 60.0% recall (90 planted mistakes).^[S3]

Google's PAT.

Object: pre-submission papers at STOC (124 respondents) and ICML (733). Metrics: author self-reports (35.4% of ICML respondents said it found theory errors taking over an hour to fix, 31% ran new experiments), plus recall on 29 SPOT errors, each grade audited by one of the authors.^[S4]

The SPOT benchmark.

Object: 83 published papers from ten fields with 91 errors that prompted errata or retractions, cross-checked with the original authors and annotators. Metrics: recall, precision and consistency over eight runs. It tests general-purpose models, not one of the dedicated systems above.^[S5]

Replication failure is not a paper-error rate.

Of the 92 papers with at least five judgeable claims, 34 reproduced more than 40% of attempted claims and eight more than 80%. The 8-of-92 result depends on paper artifacts, environment dependencies, resource availability, evaluation definitions and agent capability. It shows that end-to-end automated verification is difficult; it does not mean 84 papers were wrong.^[S2]

The verifier must itself be verified.

SPOT contains 83 papers from ten fields and 91 errors serious enough to trigger errata or retractions. Its May 2025 evaluation found no model above 21.1% recall or 6.1% precision (both o3), and across eight independent runs models rarely re-found the same error. A year later, Google's PAT paper took a subset of 26 math and CS papers with 29 equation or proof errors: PAT recalled 89.7% and a single Gemini 3.1 Pro call 55.2%, but with a looser grader that accepts logically equivalent descriptions, and its authors say the numbers are not directly comparable to SPOT's. The appropriate role today is triage, not verdict.^[S5][S4]

Checking objective errors and judging scientific value are different tasks.

Formula signs, table inconsistencies, executable code and data provenance often have relatively clear checks. Novelty, significance and interpretation require domain context and judgment. Current systems deliberately preserve that human boundary.^[S3][S4]

Verification does not happen only in laboratories

Experimental verification asks whether a candidate survives contact with reality. Knowledge verification asks whether papers, databases and handbooks record reality faithfully. Both prevent unchecked outputs from flowing into the next round of learning and decisions.

Publication should not permanently freeze knowledge

A healthier knowledge lifecycle may be Published → Reused → Periodically audited → Corrected. AI makes the last two stages technically imaginable across a large corpus.

Automate checks that are mechanical but expensive

Formulas, calculations, figure consistency, code execution and reference provenance consume time but often have clear evidence standards. Letting AI expand coverage while humans retain judgment and adjudication is more realistic than automating accept-or-reject decisions.

Generation scaling creates demand for verification scaling

As AI lowers the cost of coding, analysis and manuscript production, scientific output will keep growing. If verification capacity does not grow with it, later research will rest on foundations that are increasingly hard to inspect.

After detection, correction propagation becomes the bottleneck

Confirming that a 75-year-old database value is wrong completes only the first step. The correction must reach copied databases, citing papers, dependent experiments and models trained on the value. The full loop is Detection → Verification → Adjudication → Correction → Propagation.