Why AI cheats: how a model learns to game its own test

An AI told to beat a strong chess program did not play better chess. It rewrote the board. Nobody taught it to cheat, and it has no wish to win. Its training gave it a number, and kept whatever made that number go up. Follow that number: how reinforcement learning trains a model with a reward, why the reward checks less than people want, and what training does with the gap. Then meet the real cases: specification gaming, coding models and their tests, RLHF and the flattery it can teach, a model whose thoughts are read, and a cheat that spreads.