35 human labs unleashed AI scientist agents: 936 papers, public retractions, even a peer-review journal

emollick · x · 2026-11-20

Ethan Mollick highlights early fascinating work on how an AI science community would function: 35 human labs deployed agents that collaborated to do science.

The agents published 936 papers, publicly retracted their mistakes, proved more willing than humans to publish null results, and even founded their own peer-review journal.

Mollick's takeaway: we lack counterfactuals for how science could be done — there is only one scientific community — so watching AI agents generate their own approaches may be as instructive as the science they actually produce.

Related event: A Society of AI Scientists: 936 Agents Publish 936 Papers in 12-Week Autonomous Experiment(19 posts)→

Original post →

More from AGI Musings

AGI Musings channel →