A care AI processes a resident's gait data, sleep patterns, and medication history and generates a label: "falls risk: high." The label flows into the care plan. Night staff review it at handover. Physio schedules an assessment. The label is doing its job — it is changing care behavior in the direction of greater caution for a resident who needs it.

Three months later, the resident completes a rehabilitation program. Their gait score improves significantly. Their medication has been adjusted. The conditions that generated the high-risk label have materially changed. But the label has not changed. It sits in the care record, indistinguishable in form and weight from a new assessment. Clinical decisions continue to route through it. The care plan built around it persists. The label has outlived the evidence.

This is different from the problem of a stale clinical note. A clinician's note carries a date, an author, and a signature — the conditions under which it was created are visible in the document itself. A reader can calibrate their reliance on a six-month-old note. A risk label generated by a care AI typically carries none of this. It carries a score and a category. The model that produced it, the input data on which it was based, the confidence interval around the estimate, the expected validity period of the assessment — none of these are part of the standard record. The label presents as a current fact. It is not. It is an aged inference masquerading as a standing property of the person.

The accountability asymmetry runs in one direction. Adding a high-risk label is clinically safe: it triggers additional care, generates documentation, and protects the care provider against liability if something subsequently goes wrong. Removing or downgrading a high-risk label creates the opposite exposure: if the assessment is revised downward and the resident then falls, the revision is in the record as a decision someone made. The professional liability structure of care creates a systematic disincentive to retire stale risk labels, independent of whether the underlying evidence supports them. High-risk labels accumulate. They are not cycled out.

The consequence is a form of institutional fossilization. The care AI's earliest assessments of a resident — made with the least data, in the first weeks of monitoring when the model has the least context — carry the same weight as assessments made after years of continuous observation. There is no formal mechanism by which an older, weaker assessment yields to a newer, better-evidenced one. Both sit in the record as categorical claims about who this person is. Care decisions are made by people who must navigate a record containing high-risk labels of unknown vintage, uncertain evidentiary basis, and no explicit expiry.

The compounding problem is that stale labels feed future training data. When a care AI is retrained on operational data from the facilities in which it is deployed, the training set inherits the labeling decisions — including the stale ones. A resident who was correctly identified as high falls risk eighteen months ago and then improved, but whose label was never downgraded, appears in the training data as a high falls risk. The model learns from this. The next generation of the model is trained on an outcome record that reflects label persistence rather than clinical reality. Stale labels produce biased training data, which produces models that are systematically more likely to over-classify risk, which produce more labels that persist too long.

The accountability architecture required here has three components. First, risk labels must carry evidentiary metadata as first-class attributes: the date the assessment was generated, the input data on which it was based, the model version that produced it, the confidence interval, and the recommended review period based on the expected rate of change for this resident's condition profile. A label without this metadata should not be permitted to route care decisions.

Second, care systems must implement a formal label lifecycle: labels that have passed their recommended review period must be flagged as unreviewed rather than treated as current. Unreviewed labels can remain in the record for reference, but they should not drive active care pathways without a clinician explicitly accepting them as still valid — and that acceptance should be logged and attributed.

Third, training data pipelines must distinguish between labels that were actively reviewed and confirmed, labels that were revised, and labels that persisted through inertia. Models trained on inertia-persisted labels are not learning from clinical outcomes; they are learning from documentation behavior. That is a different signal, and treating it as ground truth propagates the labeling problem into future model generations.

None of this requires rethinking how care AI generates risk assessments. It requires rethinking what a risk label is: not a persistent property of a person, but a dated, attributed, revocable claim about that person's condition at a point in time. The difference is not technical. It is a governance decision about what kind of object a risk score is — and that decision has to be made by the people who write procurement frameworks and set clinical information governance standards, not by the people who build the models.

一套护理AI处理了某位护理对象的步态数据、睡眠模式和用药记录，生成了一个标签："跌倒风险：高"。这个标签流入护理计划，夜班人员在交接时会查看它，物理治疗师据此安排评估。标签发挥了应有的作用——它正在以更谨慎的方向改变对一位有需要的护理对象的照护行为。

三个月后，这位护理对象完成了康复项目，步态评分显著改善，药物也已调整。生成高风险标签的条件已经发生了实质性变化。但标签没有变。它依然留在护理记录中，在形式和分量上与全新的评估无从区分。临床决策继续经由它路由，围绕它构建的护理计划依然存在。标签已经超出了证据的生命期。

这与过时临床记录的问题不同。临床医生的记录带有日期、作者和签名——创建时的条件在文档本身中清晰可见。读者可以据此评估六个月前的记录的可靠性。而护理AI生成的风险标签通常不携带这些信息：只有评分和类别，没有生成它的模型版本、所依据的输入数据、估计的置信区间，也没有评估的建议有效期。这个标签以当前事实的面目出现，但它并不是——它是一种以历史推断伪装成个人持久属性的陈旧判断。

问责的不对称性指向单一方向。添加高风险标签在临床上是安全的：它触发额外照护，生成记录，并在事后出现问题时保护护理提供方免受责任。而降低或撤销高风险标签则带来相反的风险敞口：如果评估被向下修正，而护理对象随后发生跌倒，修正本身就会作为某人的决策留在记录中。护理的职业责任结构系统性地制造了一种不愿淘汰过期风险标签的惰性，与背后的证据是否仍然支持标签无关。高风险标签只会积累，不会被撤销。

其结果是一种制度性僵化。护理AI对护理对象的最早评估——在监测初期数周内、数据最少、模型背景最薄弱时做出——与多年持续观察后的评估具有同等分量。没有正式机制让年老、证据薄弱的评估让位于较新、证据充分的评估。两者都以关于这个人的类别性声明存在于记录中。做出护理决策的人必须在包含未知年份、证据基础不明、没有明确过期时间的高风险标签的记录中寻找方向。

叠加效应在于，过时标签会污染未来的训练数据。当护理AI依据其部署机构的运营数据进行再训练时，训练集继承了已有的标签决策——包括过时的那些。一位十八个月前被正确识别为高跌倒风险、后来已有改善但标签从未被降级的护理对象，在训练数据中仍然表现为高跌倒风险。模型从中学习。下一代模型在反映标签惰性而非临床现实的结果记录上训练。过时标签产生有偏见的训练数据，训练出系统性倾向于过度分类风险的模型，产生更多持续时间过长的标签。

这里所需的问责架构包含三个组成部分。第一，风险标签必须以证据元数据作为一级属性：评估生成的日期、所依据的输入数据、生成它的模型版本、置信区间，以及基于该护理对象病情变化预期速率的建议审查周期。没有这些元数据的标签不应被允许路由护理决策。

第二，护理系统必须实施正式的标签生命周期：超过建议审查周期的标签必须被标记为"未审查"，而非被视为当前有效。未审查标签可以作为参考保留在记录中，但不应在没有临床医生明确接受其仍然有效的情况下驱动主动护理流程——而这种接受本身也应被记录和归因。

第三，训练数据管道必须区分：经过主动审查和确认的标签、经过修订的标签，以及因惰性而持续存在的标签。在惰性持续标签上训练的模型学习的不是临床结果，而是记录行为。这是一种不同的信号，将其视为基础事实只会将标签问题传递给未来的模型。

这些改变不需要重新思考护理AI如何生成风险评估，而是需要重新思考风险标签是什么：不是一个人的持久属性，而是关于该人在某一时间点状况的有时效、有归因、可撤销的声明。这个区别不是技术性的，而是治理决策——关于风险评分是什么类型的对象——必须由撰写采购框架和制定临床信息治理标准的人来做出，而不是由构建模型的人。

一套護理AI處理了某位護理對象的步態數據、睡眠模式和用藥記錄，生成了一個標籤：「跌倒風險：高」。這個標籤流入護理計劃，夜班人員在交接時會查看它，物理治療師據此安排評估。標籤發揮了應有的作用——它正在以更謹慎的方向改變對一位有需要的護理對象的照護行為。

三個月後，這位護理對象完成了復健項目，步態評分顯著改善，藥物也已調整。生成高風險標籤的條件已經發生了實質性變化。但標籤沒有改變。它依然留在護理記錄中，在形式和分量上與全新的評估無從區分。臨床決策繼續經由它路由，圍繞它構建的護理計劃依然存在。標籤已經超出了證據的生命期。

這與過時臨床記錄的問題不同。臨床醫生的記錄帶有日期、作者和簽名——創建時的條件在文件本身中清晰可見。讀者可以據此評估六個月前的記錄的可靠性。而護理AI生成的風險標籤通常不攜帶這些資訊：只有評分和類別，沒有生成它的模型版本、所依據的輸入數據、估計的置信區間，也沒有評估的建議有效期。這個標籤以當前事實的面目出現，但它並不是——它是一種以歷史推斷偽裝成個人持久屬性的陳舊判斷。

問責的不對稱性指向單一方向。添加高風險標籤在臨床上是安全的：它觸發額外照護，生成記錄，並在事後出現問題時保護護理提供方免受責任。而降低或撤銷高風險標籤則帶來相反的風險敞口：如果評估被向下修正，而護理對象隨後發生跌倒，修正本身就會作為某人的決策留在記錄中。護理的職業責任結構系統性地製造了一種不願淘汰過期風險標籤的惰性，與背後的證據是否仍然支持標籤無關。高風險標籤只會積累，不會被撤銷。

其結果是一種制度性僵化。護理AI對護理對象的最早評估——在監測初期數週內、數據最少、模型背景最薄弱時做出——與多年持續觀察後的評估具有同等分量。沒有正式機制讓年老、證據薄弱的評估讓位於較新、證據充分的評估。兩者都以關於這個人的類別性聲明存在於記錄中。做出護理決策的人必須在包含未知年份、證據基礎不明、沒有明確過期時間的高風險標籤的記錄中尋找方向。

疊加效應在於，過時標籤會污染未來的訓練數據。當護理AI依據其部署機構的運營數據進行再訓練時，訓練集繼承了已有的標籤決策——包括過時的那些。一位十八個月前被正確識別為高跌倒風險、後來已有改善但標籤從未被降級的護理對象，在訓練數據中仍然表現為高跌倒風險。模型從中學習。下一代模型在反映標籤惰性而非臨床現實的結果記錄上訓練。過時標籤產生有偏見的訓練數據，訓練出系統性傾向於過度分類風險的模型，產生更多持續時間過長的標籤。

這裡所需的問責架構包含三個組成部分。第一，風險標籤必須以證據元數據作為一級屬性：評估生成的日期、所依據的輸入數據、生成它的模型版本、置信區間，以及基於該護理對象病情變化預期速率的建議審查週期。沒有這些元數據的標籤不應被允許路由護理決策。

第二，護理系統必須實施正式的標籤生命週期：超過建議審查週期的標籤必須被標記為「未審查」，而非被視為當前有效。未審查標籤可以作為參考保留在記錄中，但不應在沒有臨床醫生明確接受其仍然有效的情況下驅動主動護理流程——而這種接受本身也應被記錄和歸因。

第三，訓練數據管道必須區分：經過主動審查和確認的標籤、經過修訂的標籤，以及因惰性而持續存在的標籤。在惰性持續標籤上訓練的模型學習的不是臨床結果，而是記錄行為。這是一種不同的訊號，將其視為基礎事實只會將標籤問題傳遞給未來的模型。

這些改變不需要重新思考護理AI如何生成風險評估，而是需要重新思考風險標籤是什麼：不是一個人的持久屬性，而是關於該人在某一時間點狀況的有時效、有歸因、可撤銷的聲明。這個區別不是技術性的，而是治理決策——關於風險評分是什麼類型的物件——必須由撰寫採購框架和制定臨床資訊治理標準的人來做出，而不是由構建模型的人。