Position: Do Not Explain Vision Models Without Context

Current explanation methods for vision AI are incomplete: Popular techniques for explaining how computer vision models make decisions ignore contextual relationships between objects in images—a critical flaw illustrated by the paper's example that a stethoscope's meaning depends entirely on its relationship to nearby people.

Context-blind explanations can fail in real-world applications: The paper demonstrates concrete failures of existing explanation methods and identifies real-world use cases where spatial context is essential, suggesting that current AI transparency tools may give false confidence in high-stakes domains like medical imaging or autonomous vehicles.

We need to shift from "where" to "how" explanations: Rather than only identifying which parts of an image influenced a model's decision, explanation methods should clarify how different elements interact and relate to each other—a fundamental change in approach that better reflects how humans actually understand visual reasoning.

Responsible AI deployment requires context-aware transparency: For organizations deploying vision models in sensitive applications, the paper's findings suggest that current explanation methods may be insufficient for genuine accountability, and stakeholders should demand more sophisticated explanations that account for object relationships and spatial context.

This gap represents an urgent research priority: The paper calls for new research directions to incorporate contextual information into AI explanations, indicating that the field of explainable AI for vision models is still developing and may not yet provide the transparency necessary for trustworthy AI systems.