Paper

{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.24048v1</id>\n    <title>Human Factors in Detecting AI-Generated Portraits: Age, Sex, Device, and Confidence</title>\n    <updated>2026-03-25T07:56:26Z</updated>\n    <link href='https://arxiv.org/abs/2603.24048v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.24048v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Generative AI now produces photorealistic portraits that circulate widely in social and newslike contexts. Human ability to distinguish real from synthetic faces is time-sensitive because image generators continue to improve while public familiarity with synthetic media also changes. Here, we provide a time-stamped snapshot of human ability to distinguish real from AI-generated portraits produced by models available in July 2025. In a large-scale web experiment conducted from August 2025 to January 2026, 1,664 participants aged 20-69 years (mobile n = 1,330; PC n = 334) completed a two-alternative forced-choice task (REAL vs AI). Each participant judged 20 trials sampled from a 210-image pool comprising real FFHQ photographs and AI-generated portraits from ChatGPT-4o and Imagen 3. Overall accuracy was high (mean 85.2%, median 90%) but varied across groups. PC participants outperformed mobile participants by 3.65 percentage points. Accuracy declined with age in both device cohorts and more steeply on mobile than on PC (-0.607 vs -0.230 percentage points per year). Self-rated AI-detection confidence and AI exposure were positively associated with accuracy and statistically accounted for part of the age-related decline, with confidence accounting for the larger share. In the mobile cohort, an age-related sex divergence emerged among participants in their 50s and 60s, with female participants performing worse. Trial-level reaction-time models showed that correct AI judgments were faster than correct real judgments, whereas incorrect AI judgments were slower than incorrect real judgments. ChatGPT-4o portraits were harder and slower to classify than Imagen 3 portraits and were associated with a steeper age-related decline in performance. These findings frame AI portrait detection as a human-factors problem shaped by age, sex, device context, and confidence, not image realism alone.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.HC'/>\n    <published>2026-03-25T07:56:26Z</published>\n    <arxiv:comment>36 pages, 15 figures, 1 supplementary table. Project page: https://github.com/gdrpaul3-byte/hsmu_ai_detection_public</arxiv:comment>\n    <arxiv:primary_category term='cs.HC'/>\n    <author>\n      <name>Sunwhi Kim</name>\n      <arxiv:affiliation>Hwasung Medi-Science University, Dept. of Bio-Healthcare, South Korea</arxiv:affiliation>\n    </author>\n    <author>\n      <name>Sunyul Kim</name>\n      <arxiv:affiliation>Yonsei University, Graduate School of Engineering, Dept. of Artificial Intelligence, South Korea</arxiv:affiliation>\n    </author>\n  </entry>"
}