LLMs can de-anonymize authors, threatening double-blind peer review

Double-blind peer review is a cornerstone of scientific integrity, designed to prevent status and affiliation bias by keeping authors anonymous to reviewers. But a new arXiv preprint (arXiv:2608.05157) by Bulambo Mwendelwa Gloire and Prasenjit Mitra warns this process is increasingly fragile in the age of large language models. The researchers demonstrated that LLMs can collapse anonymity efficiently, using only titles and abstracts to narrow down authorship to a small subset of plausible candidates from a pool of five domain experts. Critically, the model's success persisted even when stylistic and bibliographic cues were stripped away, indicating that topical framing and problem selection create a "latent conceptual signature" for each researcher.

The implications are profound. If LLMs can reliably infer authors from semantic content alone, the foundational assumption of double-blind review—that anonymized manuscripts convey merit without revealing identity—breaks down. The authors argue this vulnerability necessitates rethinking how anonymity and fairness work in an AI-augmented research ecosystem. Rather than simply banning LLMs from review processes, the paper suggests new protocols may be needed, such as AI-resistant anonymization, adversarial rewriting, or shifting toward transparent post-publication review. While the experiment was limited to abstract-level inference, the threat is real: as LLMs improve and gain access to broader literature, author identification could become trivial, undermining equity in scientific publishing. The paper adds to a growing body of evidence that AI is reshaping not just how research is done, but also how it is evaluated, and demand urgent attention from journals, funding agencies, and academic institutions.