Beyond forced choice: hyper-realistic AI faces reveal a four-cluster perceptual geometry of emotion

Forced-choice paradigms in facial emotion perception research require observers to commit to a single label and discard any simultaneous activation of related categories that participants might otherwise express. We introduce a multi-select rating paradigm in which 112 participants rated each of 52 facial-expression stimuli on 13 emotion categories using continuous 0-to-10 sliders, alongside a separate authenticity slider. Stimuli were generated using OpenAIs DALL-E 3 from an orthogonal four-factor design (emotion, race, gender, age), allowing demographic balance and compositional uniformity more readily than photograph-based stimulus sets permit. Linear mixed-effects models with participant random intercepts and logistic regression with cluster-robust standard errors were applied to 5,824 trials. Perceived authenticity reliably predicted agreement between the participants top-rated category and the AI-prompted target emotion (odds ratio 1.03 per +1 unit, p < .001), and a principal-components analysis of the 13-emotion rating matrix recovered three interpretable dimensions accounting for over half of trial-level variance, with the first two corresponding to the canonical valence and arousal axes. Hierarchical clustering exposed a tight cluster of self-conscious negative emotions (sadness, embarrassment, shame) that does not align with the canonical basic-versus-complex emotion distinction. Collapsing the participants top-rated emotion to its cluster raised the trial-level prompt-agreement rate from 43.1% to 86.9% ({Delta} = +43.8 percentage points), indicating that participants most common departures from the prompted target fell within rather than across clusters. Among Autism-Spectrum Quotient subscales, the Social-Skill subscale showed a robust interaction with contempt-target faces (p < .001), with higher Social-Skill scores predicting elevated contempt ratings of approximately 1.2 points on the 0-to-10 scale. Two further secondary patterns emerged in the AQ-subscale-by-emotion grid, including an opposing-direction effect of Imagination and Attention-to-Detail subscales on awe ratings. We acknowledge that some of the recovered structure may partly reflect statistical regularities of the AI-generation pipeline rather than human perceptual geometry alone, and we discuss this caveat at length below. Multi-select rating with hyper-realistic AI-generated stimuli produces a richer perceptual probe than the forced-choice paradigms that have dominated this literature, and reveals subscale-level autistic-trait associations that AQ-total analyses obscure.