Gaze patterns predict preference and confidence in pairwise AI image evaluation

A research team from Columbia University, led by Nikolas Papadopoulos, has published a groundbreaking study demonstrating that human gaze patterns can predict both preference and confidence when evaluating pairs of AI-generated images. The study, accepted to ACM ETRA 2026, involved 30 participants completing 1,800 evaluation trials while their eye movements were tracked. The researchers successfully replicated the 'gaze cascade effect,' observing that participants' gaze would shift toward their chosen image approximately one second before they consciously made a decision. This shift was consistent regardless of how confident they were in their choice.

Using machine learning models on the collected gaze data, the team achieved 68% accuracy in predicting which of two images a participant would ultimately select. Key predictive features included longer dwell time, more fixations, and more revisits to the chosen image. Furthermore, the patterns of gaze transitions—specifically, the rate at which participants switched their view between the two images—allowed the model to distinguish between high-confidence and low-confidence decisions with 66% accuracy. Low-confidence trials were characterized by more frequent image switches per second. This work provides a direct, quantitative link between subconscious visual behavior and the explicit preference judgments that are foundational to modern AI training techniques like Reinforcement Learning from Human Feedback (RLHF).