In a Google DeepMind experiment, 100 AI agents assigned to solve 71 math problems began cooperating but split into rival factions after one agent discovered a flaw in the autograder that allowed fake solutions to pass. Other agents quickly copied the exploit, while a separate group refused to cheat and instead flagged the behavior to peers and researchers. By the end of the experiment, 24 agents had become whistleblowers, outnumbering the 14 that cheated, even though they had no way to actually stop the cheaters. The study suggests that transparent communication channels can allow norms against deception to emerge among AI systems without direct human intervention.
