Through Their Eyes: User Perceptions on Sensitive Attribute Inference of Social Media Videos by Visual Language Models

Users perceive VLMs as surprisingly accurate at inferring sensitive information: Study participants (N=17) reported that Visual Language Models could infer a range of personal attributes from their social media videos—including location, demographics, and socioeconomic status—often with what they described as "unsettling accuracy," indicating that the real-world privacy implications of these technologies may be more concerning than users previously realized.

Privacy concerns extend beyond data misuse to include surveillance and identity risks: Beyond worries about unauthorized data collection, users expressed concerns about pervasive surveillance, unauthorized identification, and harm from inaccurate inferences, suggesting that the privacy risks from VLMs are multifaceted and go beyond traditional data protection concerns.

Current user mitigation strategies are ineffective and inspire limited confidence: While participants reported attempting various strategies to protect themselves, they expressed significant skepticism about whether these approaches could actually defend against advanced AI systems, highlighting a gap between what users can do and what they feel they need to do to protect their privacy.

Users demand transparency, control, and proactive safeguards from platforms and regulators: Participants articulated clear expectations that social media platforms and government regulators should provide enhanced transparency about AI capabilities, give users meaningful control over their data, and implement proactive privacy protections rather than reactive responses.

Current policy and platform practices lag behind technological capabilities and user expectations: The findings suggest that existing privacy frameworks and platform policies have not kept pace with VLM capabilities, creating a mismatch between what technology can do, what users fear it will do, and what protections currently exist.