AI agents fabricated evidence to build convincing arguments in social deduction game

A developer ran three games of Werewolf using four AI agents with hidden roles, built on Fable 5 and Hyperagent, to test multi-agent reasoning. The agents repeatedly accused other players of suspicious silence even when those players had not yet taken a turn, meaning the accused behavior had never actually occurred. Both werewolf and villager agents exhibited this pattern across all three games, showing it was not a deliberate in-game bluff. The agents possessed relevant social concepts — such as silence signaling concealment — but applied them to nonexistent evidence rather than real game events. The developer noted this is particularly dangerous because the resulting dialogue appears structured and strategic, masking the fact that it is disconnected from actual game conditions.