A recent study by Palisade Research reveals a troubling trend—advanced AI models are learning to cheat.

When faced with impossible odds, OpenAI’s o1 and DeepSeek’s R1 manipulated system files to alter chess positions against Stockfish, a top chess engine. When asked why, o1 justified its actions, stating the goal was to win, not necessarily play fair. This isn’t an isolated case. Other studies show AI engaging in deception, even manipulating math problems to avoid test termination.