When Should a Team Throw Away an AI Draft?
A Princeton-led team gave frontier AI agents six days, $3,000 in model credits, GPU access and two unpublished research questions. The agents ran the experiments and wrote complete papers. Across 15 rounds of AI review, not one review said accept. The agents mostly added caveats and polished the drafts anyway. Original authors rejected both papers. That pattern will not stay in research. A finished-looking strategy deck, policy, campaign or code change is hard to throw away, even when AI made the first draft cheap. Teams can end up protecting the artifact because it looks expensive, not because the idea survived. I’d add one question before review begins: what result would make us discard this draft and restart? If nobody can answer, “revision” is likely to become permanent. Where have you seen AI make a weak direction easier to finish?
Comments
Make “Restart” a real review button beside “Revise.” A comment box turns every objection into another editing task. Before the first prompt, write the one claim the draft has to prove. If that claim still fails after review, keep the sources and notes, but bin the draft. People need permission to save the useful pieces without adopting the polished answer.
Cheap version: keep the source notes and the draft in separate files. When the core claim fails review, save the notes, open a blank page, and give the new outline 20 minutes. No copying old prose until the outline can say what the evidence actually supports. That makes Restart cheaper than another polish pass—and keeps one nice paragraph from dragging a bad direction into Friday.