Can an AI reviewer catch deep flaws in research papers that expert human reviewers let through, and where does it stop?

This explores whether machine reviewers can find the serious errors (broken proofs, unsound experiments) that slip past expert human reviewers, and where that ability stops. The short answer from the corpus is yes, in a narrow but important way. The way it succeeds also explains why it can't replace human reviewers.

The strongest evidence comes from an agentic reviewer that spends extra compute at review time checking proofs and experiments line by line, instead of reading the paper once and giving a verdict Can inference scaling help reviewers catch errors humans miss?. It caught 34% more mathematical errors than one-pass AI review, and it found critical flaws in papers already accepted at STOC and ICML, two venues with demanding expert review. The key is the method more than the model. Human reviewers rarely have time to re-derive every step of a proof, and a patient agent does. The same pattern appears in evaluation more broadly: an 'agent-as-a-judge' that actively gathers evidence before ruling was about 100 times more consistent than a model judging in one pass Can agents evaluate AI outputs more reliably than language models?. That study also found a warning sign: errors in the agent's memory module spread through the rest of its judgments. Systems that dig deeper also have more places to go wrong.

The less obvious part is that being good at finding flaws is different from being a good reviewer. A study of AI reviewers found a 'hivemind' effect: different AI systems agree with each other far more than human reviewers do, so the field would lose the range of perspectives that peer review relies on Can AI systems safely replace human peer reviewers?. Worse, simply rewording a paper's text, with no change to the science, raised AI scores by almost half a point. An AI that finds a hidden error in a proof can still be swayed by polished prose. Human reviewers have a matching blind spot. One of three fully AI-generated papers passed double-blind review at an ICLR workshop, and its own creators later found a citation error and judged none of the three good enough for the main conference Can AI-generated papers pass peer review undetected?. Fluent, confident writing can hide errors from both kinds of reader, a pattern that also shows up in deployed AI, where strong overall accuracy hides rare, costly mistakes Why do confident wrong answers hide in standard accuracy metrics?.

So the most promising designs don't swap one reviewer for the other. They split the work. In a randomized trial at ICLR 2025, AI feedback on human-written reviews led 27% of reviewers to revise them, and blinded raters judged the revisions more specific and informative Can LLM feedback help peer reviewers improve their own reviews?. On the paper-writing side, one system separates the parts that need a model's judgment from the parts that can be checked mechanically, which shrinks the area where errors can hide Can separating judgment from verification improve research paper reliability?. Proposals for AI-generated research venues use automated review-and-revise loops with defenses against prompt injection, which suggests their designers already expect gaming Can automated review loops handle AI-generated research at scale?.

The unexpected takeaway is that machine reviewers are strongest where review is closest to verification, such as checking each step of a proof or rerunning an experiment, and weakest where review is judgment, such as deciding whether work matters or resisting persuasive framing. Human expert review often fails at the first because of time and fatigue, so combining the two covers each one's weaknesses. The corpus doesn't yet show how often machine-found flaws in real reviewing hold up under expert checking, so treat the STOC/ICML results as promising rather than settled.