Can "bad moves" actually beat AI at Go? Ke Jie reveals new tactics

Go world champion Ke Jie recently revealed a novel anti-AI tactic during a livestream: deliberately playing “blunders” that violate fundamental go principles—such as crawling along the first or second lines or filling in one’s own eyes—to create board positions rarely seen in AI training data, thereby tricking the engine into misjudging the game state. This approach has already been demonstrated in practice: Japanese player Ryunosuke Shibano has consistently defeated top-tier go AIs while receiving a nine-stone handicap, and the FAR AI research team confirmed its effectiveness through controlled experiments back in 2022.

What Ke Jie describes—jokingly calling it “acting dumb to bait the AI”—isn’t a victory of skill but a precise strike at an algorithmic blind spot. The tactic exploits a core assumption in most go AIs: that opponents play rationally. When humans instead feed the model absurd moves, the AI often misreads its own position as overwhelmingly strong and overlooks critical weaknesses, leading to collapse.

That said, the tactic isn’t foolproof. As AI models evolve, many early vulnerabilities have been patched, narrowing the window where such strategies work. Ke himself stresses this is a niche, non-orthodox method—useful only in specific contexts, not a blueprint for competitive go. Yet from a game theory perspective, it embodies the logic of mixed-strategy Nash equilibrium in imperfect-information settings: when your opponent assumes rationality, calculated irrationality can become optimal.

FAR AI’s 2022 experiments already showed that sustained, deliberate deviation from standard play can reliably destabilize leading go engines. But with newer versions closing these gaps, the window for such exploits is shrinking. If you test this against older AI builds like KataGo or Leela Zero, you may still see breakdowns—but remember Ke Jie’s caveat: this targets model flaws, not go itself.