What happened
A preprint posted Nov. 19 and reported by CryptoSlate on Saturday found that Anthropic's Claude correctly identified the author of a peer-review report 56% of the time across a dataset of 398 reports contributed by six volunteer researchers. The experiment compared stylistic, lexical, and structural fingerprints in the anonymous reviews against each volunteer's prior public writing.
A coin flip would have scored near the base rate for a six-author pool, roughly 17%. Claude more than tripled that. The authors framed the result as a proof of concept rather than a production attack, but the methodology is reproducible with any frontier model and a modest corpus of a reviewer's past work.
The dataset covered multiple disciplines, and the paper did not publish the individual reviewers' names.
Why it matters
Double-blind peer review is the scaffolding under modern science. It's also under pressure. The preprint's finding doesn't mean anonymity is dead, but it means a determined party with API access and a target's publication history can now put a name on a review with better-than-coin-flip odds.
That changes the threat model. Retaliation, coordinated gaming of acceptance decisions, and reputational attacks on specific referees all get easier. For crypto, the stakes are concrete.
Protocol security papers, cryptography research, and economic-mechanism design pieces routinely go through blind review before publication. If reviewers suspect they can be unmasked, the honest-but-harsh review, the one that catches a flaw in a zero-knowledge proof or a consensus assumption, gets harder to write. That's a tax on research quality the field cannot afford.
