Researchers watched chess players' eyes to catch the moment they stop trusting their AI teammate.
The study used "hand and brain" chess, a variant where teammates split every move: one player picks the piece type, the other decides which piece of that type actually moves. Twenty-one chess players teamed up with an AI partner across more than 1,100 turns, choosing each time whether to keep their current role or trade with the AI. Most players kept whatever role they'd had on the previous turn. When they did switch, it showed up first in their eyes - more exploratory gaze patterns - and the move that followed tended to be worse.
The researchers built a classifier from gaze and game-state signals to predict a switch before it happened. It beat a random guess (a PR-AUC of 0.56 versus a 0.39 baseline) but not by much. The more interesting result came from interviews: players said choosing the piece type felt like more control, even though both roles required handing half of every decision to the AI regardless.
That gap between what "control" means on paper and what it feels like to a person is the actual finding here, and it is a small, unglamorous one. If a lab needs eye-tracking gear and a chessboard just to catch a faint signal of when someone wants their AI to step back, "ask the user" remains the more reliable interface for now.