NEJM AI runs an invitation-only Fast Track process that combines human and AI review and promises an initial decision within seven days. Invitations go only to manuscripts that multiple editors judge likely to be accepted. A human editor first writes a review without AI input, and then uses two large language models, GPT-5 with Thinking and Gemini 2.5 Pro, to generate structured reviews covering a summary, major and minor comments and an overall assessment. A statistical editor works with a model through a structured six-step conversation to carry out a statistical review. An editorial by Arjun K. Manrai, David Ouyang, Joseph W. Hogan and Isaac S. Kohane describes the process, which was used for the first two articles in its issue, both randomised controlled trials of ambient AI scribes.
Benefits & outcomes
- The editorial reports that the human and AI reviews contained “both overlapping and complementary critiques”, and that the published articles were revised substantively in response to both.
- The editorial reports that the AI was highly effective at checking adherence to reporting guidelines and alignment between the reported analyses and the Statistical Analysis Plan.
Notes & quotes
- An early unstructured AI statistical review “generated multiple unsolicited critiques and suggestions that […] were sometimes off topic”, and the six-step process was introduced “to establish GPT-5 as an assistant rather than a reviewer.”
- “We emphasize the central role of human judgment and values while using AI to accelerate science.” The editorial authors.
- Clarke & Esposito, in The Brief, discuss the process alongside the Purpose-Led Publishing pilot, and conclude that AI helps with structured tasks and is limited for interpretive judgment.
Publisher / company: GoogleNEJM GroupOpenAI
Date of case study: 2025
Status: Live
Built by: Publisher or platform
Sources: