We are happy to share our new open-access paper, “Emotional and attentional biases of AI-generated visual stimuli: An eye-tracking study.” We asked whether a diffusion-based generative AI model can be used to systematically manipulate the emotional content of images, and how people respond to those images in terms of self-reported emotion, visual attention, and pupil dilation.
Published in: Computers in Human Behavior Reports
Authors: Shahab Fatemi, Vahid Shalchyan
Read the paper (open access): https://doi.org/10.1016/j.chbr.2026.101394
What we did
- Took 24 neutral object images (e.g., a shirt, a cup, a lamp) and used the MagicMix diffusion model to blend each with the prompt “happy face”, creating 24 emotionally biased counterparts.
- Matched the two sets for color, brightness, contrast, and visual complexity.
- Recorded gaze and pupil size from 44 participants with an EyeLink 1000 Plus eye tracker (1000 Hz) during free viewing.
- Collected valence and arousal ratings with the Self-Assessment Manikin (SAM).
Key findings
- Self-reported emotion: Emotional images were rated significantly higher in both valence and arousal than their neutral counterparts.
- Attention: Participants looked longer at the emotional images (median dwell time of 4.68 s vs. 3.77 s).
- Pupil dilation: Emotional images produced larger baseline-corrected pupil dilation, significant in the 1.0–2.5 s window and in the trial-level model for 2.5–6.0 s.
Important caveats
The emotional condition is confounded with the presence of smiling faces and anthropomorphic features. The effects should therefore be read as specific to this type of AI-generated manipulation, not to AI-generated emotional content in general. The study also covered only positive emotion, and the AI model served solely as a stimulus-generation tool, not as an interactive agent that influences people.
Why it matters
Generative AI looks like a promising, flexible tool for designing controlled stimuli in neuroscience, human–computer interaction, and experimental design. Future work should add faceless emotional controls, negative-valence stimuli, and complementary measures such as EEG or skin conductance.
Cite this paper
Fatemi, S. S., & Shalchyan, V. (2026). Emotional and attentional biases of AI-generated visual stimuli: An eye-tracking study. Computers in Human Behavior Reports, 24, 101394. https://doi.org/10.1016/j.chbr.2026.101394
