A group of researchers has outlined a five-stage framework that would run submissions through locally-hosted AI models before any human reviewer sees them, though it exists only as a proposal for now.
The design, developed for a health sciences journal, routes a manuscript through sanitization and anonymization, parallel AI pre-screening by three separate open-weight models, an automated check gate, blinded human review, and finally editorial adjudication, with authors able to appeal at two of those stages. Because the AI reviewers run on infrastructure the journal controls rather than third-party services, the authors say the setup sidesteps the confidentiality concerns behind NIH and NSF bans on feeding unpublished proposals to outside generative AI tools. Human reviewers keep final say throughout. The paper frames this as a transparent alternative to what's already happening quietly across the field.
That quiet reality is stark: an independent analysis found roughly 21 percent of ICLR 2026's 75,800 peer reviews were fully AI-generated, up from 15.8 percent involvement in 2024, alongside documented cases of hallucinated citations and prompt-injection text hidden in manuscripts to game AI reviewers. With NeurIPS 2025 pulling in over 21,000 submissions and reviewer supply not keeping pace, the pressure pushing labs toward undisclosed AI use isn't going away on its own.
Worth noting: the same paper cites prior work benchmarking open-source LLMs on manuscript classification at just 35 percent exact-match accuracy, which is exactly why the authors kept humans in charge of every actual decision. A sensible guardrail for a system that, so far, is still a proposal rather than something running in production.