Not exactly the same thing, but I tried to use two AI models (ChatGPT 5.2 and the latest Gemini) to serve as ersatz Referees for an applied mathematics paper I am planning to publish, and it was an exercise in pointless, frustrating disaster. Suggested extensions that made no sense, requests for intermediate steps that then they couldn’t make any sense of, suggestions to introduce lemmas and remarks that were nonsensical, all the way to recommendations to state that I had proved the exact opposite of what I actually had proved. Never again.
