Revolutionary Go Strategy Outsmarts Elite AI but Falls Short Against Amateur Players

Researchers demonstrated that KataGo, an advanced Go-playing AI, can be tricked into a premature defeat using an adversarial strategy. This strategy manipulates KataGo’s overconfidence by convincing it that it has secured a winning position, despite the presence of opposing stones that nullify its territory’s points. Consequently, KataGo passes its turn, mistakenly believing the game is won, which allows the adversary to also pass and end the game. Although the adversarial strategy effectively exposes a flaw in KataGo, it is not highly skilled at Go, indicating that even strong AI systems can possess vulnerabilities. This research highlights that AI, while performant on average, may fail in unexpected ways, posing risks in critical systems like self-driving cars. Gleave emphasizes the importance of enhanced testing protocols for AI to identify these vulnerabilities, underscoring the broader implications for safety in various applications.

Source link