Clinical decision-making rarely begins with all the facts. Doctors must decide not only what additional information to collect, but also when the available details are enough to act. That moment of judgment—whether the information at hand is sufficient—is what the authors call determinability.
A new paper on arXiv probes how large language models handle this challenge in medicine. The researchers highlight a specific concern they term 'factor coverage bias': the possibility that LLMs may judge a situation as determinate based on which factors are present, rather than whether the information truly supports a decision.
The work points to a subtle but critical risk in AI-assisted healthcare. If models cannot reliably judge when to stop asking, they may either push for unnecessary tests or prematurely settle on a conclusion. The abstract does not yet provide results, but the framing suggests a need for careful evaluation of LLM reasoning under underspecified conditions.