Google’s DeepMind AIs, like AlphaZero, mastered complex games such as chess and Go by playing themselves. Yet they falter against simple games relying on mathematical logic-like Nim, where players remove matchsticks from a pyramid until one remains.
Nim belongs to a class of 'impartial' games where both players share the same pieces and rules. Winning depends on evaluating a parity function-a math operation that determines optimal moves. While humans grasp this quickly, AI trained via self-play fails to learn it.
Researchers Bei Zhou and Soren Riis tested an AI on seven-row Nim. After 500 training rounds, its performance plateaued. Even random move selection performed as well. The AI couldn’t distinguish winning moves from losing ones, indicating a systemic failure in learning abstract mathematical rules.
The flaw isn’t just in Nim. The study suggests similar blind spots exist in chess-playing AIs, which sometimes miss mating sequences despite high initial evaluations. These failures point to a core limitation: Alpha-style training excels at pattern association but fails at symbolic reasoning.
This revelation matters for AI applications in math, finance, and strategic planning-fields where logical deduction, not just pattern recognition, is essential.