Researchers Tried Letting AI Do Science. It Failed

A multi-institution study found that today's frontier AI agents could handle the mechanics of research but failed to produce original work worthy of acceptance at a top AI conference.