Through Human Eyes: Analyzing Gaze Patterns in AI-Generated vs. Real Faces

Introduction

As generative AI technologies become more prevalent, distinguishing between real and synthetic visual content is increasingly difficult. Research shows that people not only struggle to detect AI-generated faces but may also perceive them as more trustworthy than real ones (Nightingale & Farid, 2022). But while our ability to reason out whether an image is real or synthesized is no better than a coin toss, there may be underlying biological responses that are more accurate, even if we're unaware of them. This study aimed to explore visual habits and emotional instincts we use when passively viewing faces versus actively judging their authenticity, uncovering subtle differences in how our eyes respond to the synthetic and the real.

Methodology

Ten participants with varied media familiarity took part in the study. Eye-tracking data was collected using the Pupil Invisible system. The experiment had three parts:

1. Passive Viewing: Participants viewed 28 randomized face images (14 real, 14 AI-generated) for four seconds each, without being told some were synthetic. Gaze behavior was recorded unobtrusively.

2. Active Viewing: Participants then viewed the same 28 images plus 6 new ones (2 real, 2 synthetic, 2 ambiguous). This time, they were informed that some images were AI-generated and asked to classify each as real or synthetic, with unlimited viewing time.

3. Post-Interview: Semi-structured interviews explored participants’ reasoning, attentional focus, and broader attitudes toward synthetic media.

Results

• Scanpath Similarity: There were no significant differences in overall scanpath patterns between AI and real faces in either viewing condition. Time spent per image also showed minimal variation (7.12 seconds for real vs. 7.77 for synthetic during active viewing).

• Passive vs. Active Saccade Patterns: A significant behavioral shift emerged between passive and active viewing. Horizontal saccades increased in active viewing, suggesting more deliberate scanning when authenticity was in question. Participants reported focusing on the eyes and mouth during passive viewing, but switched to finer features—skin texture, lighting, symmetry—when actively trying to detect AI.

• Fixation Differences: In the passive phase, participants fixated significantly more often on real faces than synthetic ones (t = 2.649; p = 0.046). This difference vanished in the active condition, indicating that when participants were primed to question authenticity their viewing patterns changed, suggesting potential unconscious differences between viewing real vs synthetic faces. Interestingly, images believed to be real, whether accurate or not, also attracted more fixations. However, participants rarely reached strong consensus about which faces were synthetic or real and further exploration would need to be done to strengthen this observation.

Qualitative Insights

Interviews revealed that participants rarely question image authenticity in daily life unless the context raises concerns (e.g., news or political content). Several mentioned noticing AI-generated headshots on LinkedIn or other professional sites. One participant expressed concern about the erosion of visual truth in search results and online platforms. After the study, all participants reported increased skepticism and a desire to be more critical of visual media going forward.

Next Steps

1. Improve Technical Setup: Eye-tracking consistency was impacted by glasses incompatibility and varied environments. Future work should use more accessible technology and controlled settings.

2. Expand Sample Size & Diversity: The small sample (5 with usable eye-tracking data) and limited image variety restrict generalizability. A larger, more diverse participant pool and image set (including different ages, ethnicities, and emotional expressions) would strengthen findings.

3. Investigate Perceived Realism Further: The link between perceived authenticity and fixation patterns is promising. Future studies should explore how belief in an image’s realism, regardless of accuracy, influences visual engagement.

Conclusion

This exploratory study examines how people visually and cognitively engage with AI-generated versus real faces under varying levels of awareness. Although scanpaths and viewing times did not significantly differ, participants’ eye movements and interviews showed strategic shifts, especially when prompted to judge authenticity. Real faces drew more fixations during passive viewing, but this effect disappeared when participants knew some faces could be AI-generated. Responses revealed a fragile, context-dependent sense of authenticity and a lack of consensus on what looked "real," alongside growing post-study skepticism. Despite limitations, these findings offer early insight into how we may be subtly viewing synthetic media, and show promise in further exploration into the unconscious and contextual factors that shape the way we gaze.

References

Iskra, A., & Tomc, H. G. (2016). Eye-tracking analysis of face observing and face recognition. Journal of Graphic Engineering and Design, 7(1), 5–11. https://doi.org/10.24867/jged-2016-1-005

Nightingale, S. J., & Farid, H. (2022). AI-synthesized faces are indistinguishable from real faces and more trustworthy. Proceedings of the National Academy of Sciences, 119(8). https://doi.org/10.1073/pnas.2120481119