When Visual Evidence is Ambiguous: Pareidolia as a Diagnostic Probe for Vision Models

Researchers developed a framework using face pareidolia (seeing faces in non-face objects) to test how different AI vision models handle ambiguous visual information. The study found that vision-language models like CLIP and LLaVA tend to over-interpret ambiguous patterns, while pure vision models remain more uncertain and detection models are more conservative.

• →Face pareidolia serves as an effective diagnostic tool for evaluating AI vision model behavior under visual ambiguity.
• →Vision-language models exhibit 'semantic overactivation,' systematically misinterpreting ambiguous non-human regions as human faces.
• →LLaVA-1.5-7B showed the strongest over-interpretation tendencies, especially for negative emotions.
• →Pure vision models like ViT follow an uncertainty-based approach, remaining diffuse but largely unbiased.
• →Model behavior under ambiguity is determined more by representational architecture than by score thresholds.