When comparing pairs of AI-generated images, people's eye movements reveal their eventual choice and confidence level well before they click. A new study on arXiv tracked the gaze of 30 participants across 1,800 trials and replicated the gaze cascade effect, with eyes shifting toward the chosen image roughly one second before the decision (https://arxiv.org/abs/2603.24849). Chosen images received more dwell time, fixations, and revisits, allowing gaze features to predict binary preference with 68 percent accuracy. The frequency of switches between images also distinguished high-confidence from uncertain decisions at 66 percent accuracy. The authors suggest eye-tracking can supply implicit signals that improve the quality of preference data used in training methods like RLHF and DPO.