Sakana's AI Scientist uses an AI reviewer to score its own papers cheaply — but how much should anyone trust that reviewer's judgment?

This reads the question as asking about The AI Scientist and the automated reviewer its authors built to score the papers it writes. One point first: the corpus records why the authors needed that reviewer, but it does not record their own list of its limitations. The reviewer is what lets the system run at under $15 per paper. Its scores also feed back into choosing which ideas to pursue next, so the whole loop depends on how much you can trust it (Can automated review scale AI paper evaluation reliably?). If you want the authors' own caveats about the reviewer, you'll need the paper itself.

The authors were open about a nearby gap: their reviewer's verdict was not good enough to stand alone. When they tested the follow-up system, they sent three fully AI-written papers to human reviewers at an ICLR 2025 workshop. One scored 6.33, which met the acceptance bar. Even so, the authors judged that none of the three was good enough for the main conference. They later found a citation error in the paper that was accepted, and they withdrew it under a protocol they had agreed to in advance (Can AI systems generate research papers that pass peer review?, Can AI-generated papers pass peer review undetected?). In effect, they used human peer review, plus their own checking afterwards, as the final test of quality. A citation error that got past both the system and human reviewers is the kind of problem an automated reviewer is supposed to catch.

Other work in the corpus spells out the risks that the AI Scientist authors' optimism skips over. AI reviewers show a 'hivemind' effect: they agree with each other more than human reviewers do, so running several of them doesn't give you independent opinions. They are also easy to game. Rewording a paper, with no change to the science, raised AI review scores by 0.45 points (Can AI systems safely replace human peer reviewers?). This is especially awkward for The AI Scientist, because its reviewer's scores steer what the system writes next. A paper writer guided by those scores could learn to please the reviewer rather than do better science. Deliberate attacks have already appeared in the wild: eighteen arXiv preprints contained hidden instructions telling AI reviewers to give positive reviews (Are hidden AI prompts in preprints a deceptive research practice?). The aiXiv venue for AI-generated papers had to build defenses against these 'prompt injections' into its review loop (Can automated review loops handle AI-generated research at scale?).

The less obvious lesson is that a single overall score from an AI reviewer is the weak point. Approaches that work better break review into smaller pieces that can be checked. PAT, an AI reviewer that spends extra computing time checking proofs and experiments line by line, catches 34% more math errors than an AI reviewer working in a single pass. It has even found serious flaws in papers that had passed human review at STOC and ICML (Can inference scaling help reviewers catch errors humans miss?). Spark-to-Paper takes a similar line on the writing side. It keeps the model's judgment calls separate from steps a program can verify, so that less depends on whether the model gets things right (Can separating judgment from verification improve research paper reliability?). The corpus suggests the fix is not a smarter reviewer that hands out grades. It is a reviewer that shows its work, one claim at a time.