AI safety explanations fall apart when images change — and nobody noticed until now

Researchers built a test suite to measure how stable AI explanation methods are when images get slightly altered (rotated, compressed, noise added). It turns out the explanations break down far more often than anyone measured before, especially under geometric changes, and existing tests were hiding this by not checking whether the model's predictions actually stayed the same. This matters because hospitals, autonomous vehicles, and other safety systems rely on these explanations to understand what the AI is actually looking at.

AI systems used in hospitals and self-driving cars include explanation tools designed to show humans what the model noticed before making a decision. Until now, nobody had a reliable way to test whether those explanations stay consistent when the image changes slightly — a rotation, a compression artifact, a small blur. This paper shows that without careful measurement, explanations can flip almost completely while the model's decision stays the same, which means a human trusting the explanation is being shown something misleading. The test suite forces a harder question: does the explanation actually explain the decision, or just the input? For safety-critical systems, that distinction is everything.