Trust in Vision-Language Models: Insights from a Participatory User Workshop

User trust in Vision-Language Models cannot be reliably predicted by AI-based evaluation alone. The paper emphasizes that relying on AI models as judges for validation bypasses critical insights that only come from direct engagement with actual users, suggesting that responsible deployment requires participatory research rather than algorithmic shortcuts.

Trust in VLMs is context-dependent and evolves over time. The workshop findings indicate that how users build and maintain trust in these systems is not static, implying that one-time trust assessments or generic trust metrics may be insufficient for responsible AI deployment across different use cases.

Current trust metrics for VLMs lack grounding in real user experiences. The paper identifies examining "how user trust in VLMs builds and evolves" as "an open problem," suggesting that existing frameworks for measuring trustworthiness in vision-language systems may not reflect what actually matters to people using them.

Participatory design with end-users should be prioritized over cost-cutting measures. The authors critique the trend of using AI models as substitutes for user studies, advocating that the responsible development of VLMs requires direct engagement with prospective users despite higher costs and logistical complexity.

Contextualizing trust requires understanding specific user-VLM interaction scenarios. The workshop approach highlights that trust strategies and metrics must be tailored to particular use cases rather than applied universally, emphasizing the need for domain-specific and application-specific trust frameworks.