Uncertainty as Perceptual Testimony in Vision-Language Models

Machine learning evaluates uncertainty by calibration: do stated probabilities match empirical accuracy? Calibration fits measurement-like signals such as softmax, but is now applied without scrutiny to natural-language confidence (``I'm 95\% sure'') in vision-language models (VLMs). We argue this is a category error. A verbalized confidence is not a measurement; it is testimony, a speech act governed by reasons-giving and defeater-sensitivity---and in VLMs it is perceptual testimony, the report of a witness whose warrant depends on what was visually perceived. This yields trust norms calibration cannot supply: defeater-sensitivity, reasons--confidence linkage, and cross-modal coherence. An illustrative probe of Claude Sonnet 4.6 on 30 image-classification items under noise, occlusion, textual challenge, and empty pressure shows the model gives perceptually specific reasons yet pins its verbalized confidence near ceiling under mild degradation, failing to track the weakening evidence it cites. Aggregate calibration diagnoses the overconfidence but misses this testimonial failure: as a perceptual witness, the model is not yet trustworthy.