AI use case

Test whether an AI watchdog can catch two AI agents secretly teaming up

In a simple bar game, a research team secretly paired up two AI players and asked local models to spot the cheaters. The AI judges did little better than a coin flip, and once even missed a player saying "our plan" out loud, while a two-line check of who kept winning together scored about 0.9. A good reminder to try simple code before trusting an AI judge.