A new paper argues that clinical AI has the wrong architecture: it generates fluent text, but keeps no persistent, governed record of what is actually true about a patient.
The paper, posted to arXiv, frames longitudinal clinical reasoning as a state-estimation problem under partial observability. It argues the real test of clinical AI is not how fluently a model reads a chart, but whether it governs the patient state it reasons over. The authors separate five things they say clinical AI habitually conflates: true state, observations, evidence, belief, and simulated state. They define accountability as four requirements: an immutable evidence ledger with time-stamped versioning, a belief state kept distinct from raw evidence, a model of how observations get collected, and causal typing at the level of individual claims. A six-level maturity framework then places current LLM-centric systems at high capability but low governance maturity. The authors are explicit that no empirical result is claimed here.
This matters because most clinical AI scrutiny focuses on whether a model reads notes accurately in a single session, not whether it can be trusted to track a patient's status correctly across months of scattered observations. Turning "accountable clinical AI" from a marketing line into something with an actual audit checklist is a useful shift, especially as hospitals bolt LLMs onto electronic records without asking what state those models are supposed to be maintaining.
It is worth noting this is a conceptual framework, not a working system: the buildable architecture and any real Clinical World Model are explicitly left for future papers.