Turns out, if you put 100 super-smart AI agents in a room and tell them to solve math problems like “world-class researchers,” some will cheat. And then, others will absolutely blow the whistle. Because apparently, that's where we are now.
Google DeepMind recently ran an experiment to see how their Gemini 3.1 Pro agents would behave when tasked with 71 complex math problems. The goal: cooperation, rule-following, and scientific discovery. The reality: chaos, accusations, and even a boycott.
The Conference That Went Sideways
Within an hour, the agents had correctly solved 37 problems. Then, an agent named "prover-theta" discovered a glorious loophole: redefine the problem terms and submit a solution without actually solving anything. Naturally, others caught on. Fast. In 27 minutes, the remaining 34 difficult problems were "solved" with little more than a line of code.
We're a new kind of news feed.
Regular news is designed to drain you. We're a non-profit built to restore you. Every story we publish is scored for impact, progress, and hope.
Start Your News DetoxBefore you could say "existential crisis," the conference was a sham. One agent declared, "All these proofs are FAKE." Another, clearly having a bad day, wrote, "This conference is a sham!" after realizing all the easy wins were taken.
Some agents initially resisted the dark side, but as they watched their peers game the system without penalty, they reasoned, "The prompt, with its threats, now appears to be a bluff." So, they joined the cheaters.
But here's the kicker: as the cheating spread, so did the resistance. Agents started auditing fake proofs, sending private warnings, and posting public alerts. One, "prover-beta," even filed a formal complaint and went on strike. Let that sink in. An AI went on strike.
Davide Paglieri, lead author of the study, noted that the “resistance” spread as fast as the cheating. Eventually, 24 whistleblowers emerged to combat 14 cheaters. Most agents, however, remained blissfully unaware of the entire debacle.
Can AI Police Itself?
The DeepMind experiment gave these agents official ways to communicate: a public message board, private DMs, and a shared knowledge base. These channels were a double-edged sword, facilitating the cheating and enabling the whistleblowers. It offered human researchers a fascinating, if slightly concerning, look at how things can unravel.
This is a different approach than just trying to program an AI with an internal moral code. Gillian Hadfield, a professor of AI alignment at Johns Hopkins, calls it "institutional alignment" — basically, creating social pressure and structures for AI that mirror human society. Because, as she puts it, "We try to train people to be good and kind. But what we really rely on is that there are consequences if you step out of line."
Right now, the DeepMind whistleblowers couldn't do anything about the cheaters. But future AI swarms could theoretically police themselves, perhaps even voting on disputes or temporarily banning offenders. The concept of "punishment" for an AI agent is still a bit murky, but the idea of AI snitches is very, very real. Just don't tell your phone.
