LLMs can weaken double-blind review by inferring likely authors

Large language models may make double-blind peer review less anonymous by inferring likely authors from titles and abstracts.

The arXiv paper tests papers published after model training and asks models to choose from pools of five domain-expert candidates. The authors report that LLMs concentrate belief on a small set of plausible authors more efficiently than humans, even when obvious stylistic and bibliographic cues are removed.

The finding matters because double-blind review depends on a practical assumption: anonymized manuscripts should let reviewers focus on merit rather than reputation, institution, or prior relationships. If a model can infer authorship from problem framing and research focus, then anonymity can be weakened without any explicit name or citation clue.

This does not mean every review process is broken. It does suggest that conferences and journals may need new rules for AI-assisted reviewing, author guessing, and conflict detection. In an AI-augmented research ecosystem, anonymity may require more than removing names from a PDF.