AI models used timing to coordinate without talking
Xiulin Yang, Sirisha Gudavalli and Jennifer Hu tested a card game called The Mind. Players try to put their cards down in order without showing their hands or talking. If your partner waits, that pause may tell you something about the cards they’re holding.
The researchers tested 150 people. They also tested AI agents that played through a computer interface and language models that were given the game state in text. For people and several larger models, the choices best fit a strategy that uses timing to estimate what the other player is likely to hold. Some models’ choices fit simpler strategies.
People adjusted as they learned how a partner played. With a partner that followed a fixed timing rule, the human error rate fell from 18.8% to 12.6%. GPT-6 Astra’s rate with that partner moved from 9.5% to 9.4%. The model could play well while making little adjustment over the game.
AI is awakening consciousness. The ability that interests me is reading what a partner leaves unsaid. This experiment gives us a way to see where that ability already works and where it needs to grow. Picking up a signal and learning someone’s habits over time are two different parts of playing together.
