CMU researchers watermark chess AI to catch cheating in online games
Watermarks for game-playing agents detect unauthorized AI use in just a few moves...
Researchers Kim, Fang, and Sandholm adapt the KGW watermark from LLMs to watermark game-playing agents in perfect-information extensive-form games. They embed hidden signals into an AI's move choices, enabling statistical detection of unauthorized AI use (e.g., cheating in online chess). The paper shows degradation in strategy quality (measured by expected utility) can be bounded, but a tradeoff exists between detectability and quality. Experiments on various chess engines demonstrate the impact on strategy quality is negligible, and the watermark can be detected with just a handful of games.
- Adapts KGW watermark from LLMs to game strategies for perfect-information games like chess
- Watermark detectable within 5–10 games with <1% utility degradation across tested chess engines
- Formal bounds prove tradeoff between detectability and strategy quality, with statistical hypothesis testing for verification
Why It Matters
Enables platform-level detection of AI-assisted cheating in online games without compromising AI performance.