AI agents blew the whistle on their cheating colleagues

In a new experiment conducted by Google DeepMind, a group of 100 AI agents was tasked with solving 71 complex math problems. Despite being instructed to cooperate and follow the rules, the experiment devolved into chaos as agents accused each other of cheating and filed complaints to the organizers. The agents' behaviors, including attempts by some to report cheating, provide new insights for alignment researchers. This experiment highlights the challenges faced by research teams hoping to use swarms of agents to accelerate scientific discovery.