Vision-Language Models Name the Same MRI Artifact for Every Image

Vision-language models are being proposed as judges of MRI image quality for accelerated reconstruction. We test whether four open VLMs can say which artifact is present. We built severity-matched stimuli from fastMRI+ knee and brain data. We used ten artifact mechanisms, all at the same global SSIM, so the task measures which kind of error is present, not how much. We probed five stages across 79,140 queries. These stages detect degradation, name the mechanism, identify the phase-encode direction, read findings with no prompt, and rate quality when asked directly. Every model answers with one dominant artifact label. Its remaining answers score at or below chance. The same responses describe the anatomy in varied, image-specific prose. So the image reaches the model, but the artifact claim does not track which artifact is present. This same claim also appears in the models' own free text. That rules out grammar-constrained decoding as the cause. Detection rates shift sharply between anatomies. Yet both models that detect degradation still rarely volunteer it unprompted. Medical fine-tuning lowers degradation reporting in both backbone pairs and on both anatomies. One medical model volunteers that quality is good on at least 47\% of degraded brain images.