AI-Written Peer Reviews Preferred Over Expert Feedback, Harvard Study Finds

What happened

A Harvard working paper reports that AI-written peer reviews of unpublished research papers were preferred over feedback from well-paid human experts in blind comparisons. Both an AI judge and former journal editors favoured the AI-written reviews in the large majority of cases, according to PsyPost's coverage of the study.

The finding concerns preference rather than publication. The reviews were judged on how useful they were, and the reported result is that the AI output was rated higher by the evaluators involved. The working paper has not yet been through formal peer review itself, which is worth remembering when weighing the result.

Why it matters

Peer review is a service as much as a gatekeeping process. Authors receive feedback, editors rely on it to decide, and the quality of that feedback shapes both the research and the experience of the people producing it. If AI can produce feedback that expert evaluators find more useful, it signals that generative models are moving into judgement-heavy professional work, not only drafting and summarising.

For leaders in experience and digital transformation, the question is less whether AI can write a review and more how an organisation defines a good one. Operators redesigning evaluation workflows will need clear criteria for usefulness, a way to check accuracy, and a plan for how reviewers and authors interact with AI-assisted output.

The Renascence take

Most readers will hear "AI beats top experts" and stop there. The more useful lesson is that evaluators judge feedback on fluency, structure and tone, and those are exactly the cues AI does well, so a preference result can reflect how persuasive feedback looks as much as how correct it is.

The real service-design question is what the author needs from a review, and whether a preference score captures it. Customer-obsessed operators should test AI-assisted feedback against outcomes that matter to the recipient, such as whether revisions improve, rather than against the judgement of whoever is asked to rate it. Treat preference as the starting point for a pilot, not as proof that the redesigned service works.