When people compare two AI-generated images and choose which they prefer, their eyes shift toward the chosen image approximately one second before they consciously decide. Visual attention metrics predict the choice at 68% accuracy. Gaze transition patterns distinguish high-confidence from uncertain judgments at 66% accuracy.
The decision happens in the visual system before it reaches the verbal report. This is not new in psychophysics — pre-decisional gaze bias is well-established in consumer choice research. What's new is the connection to AI training pipelines.
Current reinforcement learning from human feedback systems train on explicit preference labels — the conscious, verbal report of which option is preferred. But if the eyes reveal preference before the conscious report, then the explicit label is capturing a post-hoc rationalization of a decision the visual system already made. The label is downstream of the actual judgment.
Eye-tracking could replace or augment explicit preference labels for training AI systems. The earlier signal is potentially more honest — less contaminated by the justification process that transforms a quick visual impression into a stated preference with reasons. The reasons may be confabulated explanations for a decision that was already made by the time the person started explaining.
The practical barrier is hardware: eye-trackers are expensive and cumbersome compared to clicking a button. But the conceptual implication is immediate. If we're training AI systems on human feedback, and the feedback we collect is a noisy, delayed version of the actual preference signal, then we're systematically degrading our training data by using the convenient measurement instead of the accurate one.
The through-claim: the measurement that's easiest to collect is not always the measurement closest to the thing you're trying to measure. When the real preference signal precedes the reported preference signal by a second, the report is a proxy — and every proxy introduces noise. The question is whether the noise is random or systematic, and post-hoc rationalization is systematic.