HOW TO REFORM ECON REFEREEING WITH AI Frontier AI is clearly extremely useful in the review process, but it should be done right. The right model is not that all referees ask ChatGPT to point out strengths and weaknesses of the paper. (The point of asking human referees is to get their human judgement.) The right model is something along these lines (helpfully illustrated by Astra in the flowchart): 1. If the editor plans to send the paper out for review, they first generate a frontier AI report (e.g. using a modified version of my open source paper reviewer. I think open source is nice here because the journal can modify it to their purposes.) 2. The authors then get the AI report back from the journal and are given the opportunity to provide comments, explain how they plan to address the points in the report (if they agree with them), and push back on any points they disagree with. 3. The journal then sends out the initial manuscript, the AI-generated frontier report, and the response from the authors. This makes everything constructive and avoids cheap duplication of AI efforts as well as removing the incentives of referees to suggest comments that most good AI systems are likely to flag in any case. The next extension should be to require authors to submit a full replication package with code and data at the submission stage to check consistency between the paper, code, data, and reported results.