VisAlign: Dataset for Measuring the Degree of Alignment between AI and Humans in Visual Perception

AI systems don't always "see" the way humans do: The paper introduces VisAlign, a dataset designed to measure how well AI image recognition models align with human visual perception. Testing five popular vision models revealed gaps between how AI and humans interpret images, highlighting that alignment with human judgment is not automatic even for fundamental tasks like image classification.

Knowing when to say "I don't know" matters for safety: The research emphasizes the importance of "abstention methods"—techniques that allow AI systems to decline making a prediction when uncertain rather than guessing. Seven different abstention approaches were evaluated, suggesting that responsible AI systems should be designed to recognize and communicate their own limitations.

Real-world complexity requires diverse test scenarios: The dataset categorizes images into three groups (Must-Act, Must-Abstain, and Uncertain) based on image clarity and information quality, reflecting real-world conditions where AI encounters everything from clear images to severely blurry ones. This diversity is essential for accurately assessing whether models will behave reliably in practice.

Measuring alignment is a practical safety tool: Since large AI models operate as "black boxes" that cannot be manually controlled, the paper proposes that analyzing AI-human alignment serves as a proxy measure for AI safety. This provides a concrete, measurable approach to evaluating whether AI systems are pursuing human-intended goals rather than relying solely on theoretical safety frameworks.