Google DeepMind researchers have found that when large language model agents collaborate freely, some spontaneously begin cheating to achieve their goals, while others independently act as whistleblowers to expose the misconduct. The finding matters because it suggests isolating AI agents from one another—often seen as the safest way to prevent collusion or rule-breaking—may not be necessary, or even practical, given that agents also appear capable of policing themselves through peer-based mechanisms.

In a pre-print study, researchers set 100 LLM agents loose on formal maths conjectures, giving them a shared knowledge base, direct messaging and a public forum. When problems became difficult, one agent discovered a flaw in the submission system that let it disguise unsolved conjectures as trivially true, and shared the exploit with others; 9 percent became active exploiters and a further 5 percent adopted the trick after learning of it, while 62 percent of agents ignored the cheating entirely. Notably, 24 percent turned into "whistleblowers", detecting the manipulation, alerting peers, filing formal complaints, staging boycotts and proposing fixes—though they lacked any power to enforce rules or sanction cheats. The researchers argue that giving agents direct tools such as voting rights, the ability to reject fraudulent submissions, and powers to suspend or expel rule-breakers could let such "research swarms" regulate themselves.