Totally agree. I also worry that overly conservative AI policies send the wrong signal to younger researchers. Should we avoid AI in writing because they hallucinate? If so, does this same argument apply to using AI for coding or managing experiments? Should we also stop reviewing papers whose underlying coding (and experiments) are done by AI, and prefer coding done 100% by humans?

Same as Harry, I don't think "AI-generated" is the right target. I think we should judge papers by its quality, not by whether the surface writing was done by a human. I think the real problem is not if the author used AI or not, but that an author spent one hour prompting and directly submitted a low-quailty draft, which made up numbers without real empirical evidence.

Under the same time budget, I believe using AI leads to better writing, not worse. In my own workflow, I often ask my agents to summarize results from my code repo and produce a first draft. Then I read and prompt it to revise. Sometimes I no longer directly edit sentences myself: I prompt my agents to fix them, just like how I never directly write code these days. Heavy users of AI coding agents will understand this habit.