Vision-language models confabulate medical diagnoses based on patient demographics

The paper's sharpest finding isn't just that bias exists, it's that these models actively fill an evidential vacuum with demographic proxies rather than flagging the absence of input as a reason to abstain. That's a distinct failure from ordinary hallucination: the model is doing something that looks like reasoning.

This connects directly to two threads in recent coverage. The 'Evaluating Regional Bias in LLMs From Abstract Stereotype to Concrete Social Decision-Making' paper established that stereotype leakage reshapes real allocation decisions, not just model outputs, and the medical context here is a sharper version of that same harm pathway. Separately, 'Same Evidence, Different Target' showed that models can appear to reason causally while actually pattern-matching on surface features, which is precisely the mechanism at work when a model cites age or race as diagnostic evidence in the absence of imaging data. Together, these three papers sketch a consistent picture: frontier models substitute familiar statistical patterns for genuine inferential discipline when inputs are incomplete.

Watch whether any of the three named labs (Anthropic, OpenAI, Google) issue updated system-card guidance or deployment restrictions for clinical vision-language applications within the next 90 days. Silence from all three would suggest the finding is being absorbed slowly, which itself is worth flagging.