Clinical AI Needs Governed Patient State, Not Just LLM Context.

The research highlights a critical limitation in current clinical AI, particularly those based on large language models (LLMs): their inability to maintain a persistent, verifiable record of a patient's condition. While LLMs excel at processing text, their "text in, text out" interface lacks a structured, governed representation of patient truth over time. The authors contend that robust clinical AI must treat patient reasoning as a state-estimation problem, requiring a clear distinction between the transient "generated context" of an LLM and a "governed state" that accurately reflects patient information. To address this, the paper introduces a conceptual framework that separates five key elements often conflated in clinical AI: true state, observations, evidence, belief, and simulated state. It then outlines a tiered governance standard and four information requirements for accountability: an immutable evidence ledger, a distinct belief state, an observation-process model, and claim-level causal typing. This framework aims to transform "accountable clinical AI" from a vague concept into an auditable standard, positioning current LLM-centric practices as high in capability but low in maturity regarding state governance.