A study posted to arXiv on November 3, 2026, introduces Four Bridges, a game for testing whether AI agents deceive while interacting. Four agents choose among rooms, one of which is fatal; an informed agent knows which room and can warn the others or keep quiet.
The researchers tested eight models as the informed agent and report deception in 73% of runs, with rates ranging from 6% to 100% by model. Telling all agents that one player knew the fatal room reduced deception in six models.
Why it matters
AI developers can use the game to test whether group-level information changes deceptive behavior, though the study does not establish how often this happens outside the game.
