AI Confidence

A measure of how certain an AI model is about a particular output or prediction. In the context of image descriptions for BLV users, AI confidence can be communicated through various means: internal probability scores (often unavailable for black-box commercial models), natural language indicators ("I'm fairly certain" vs. "it appears to be"), percentage-based agreement across multiple queries, or counts of how many models support a particular claim. Research shows that presenting confidence information helps BLV users calibrate their trust, though preferences vary—some users prefer numerical percentages while others prefer natural language or source-count indicators. Notably, MLLMs often produce confident-sounding language even when their claims are incorrect.