生成式人工智能如何塑造生命终末期的最后一餐

文献信息

中文摘要

背景： 为临近死亡的人选择临终一餐常被视为一种极具个人意义的行为。与行刑前最后一餐请求这一已有充分记录的实践不同，疾病终末期的食物与餐食选择受到医疗限制、家庭动态、文化期望和伦理判断的塑造。姑息护理护理文献日益认识到，食物和用餐时间对患者及家属而言既是意义来源，也是痛苦来源。随着生成式人工智能（AI）系统日益被引入医疗情境，它们也可能在诸如临终关怀等伦理敏感领域生成建议。此类建议并非中立；它们可能反映关于安全、舒适以及何为“善终”的规范性假设。研究目的：本研究考察生成式AI如何建构并论证临终餐食建议，以及这些输出如何反映、强化或消音关于死亡的特定伦理与文化想象。研究设计：采用定性、解释性研究设计，其基础是科学技术研究（STS）以及关于非人类行动者作为规范性行动者的批判性AI伦理视角（见理论框架，第4.1节），将AI生成的输出视为文化文本而非临床指导。分析同时聚焦于所提议的餐食及伴随这些餐食的论证性语言。参与者与研究情境：不涉及人类参与者。数据集由单一大型语言模型生成的75份AI餐食提案组成。十五个系统变化的提示各生成五次，操纵了年龄、文化背景、宗教取向、家人在场、吞咽困难和偏好偏离。伦理考量：作者所在机构的研究伦理委员会确认，该研究因不涉及人类参与者、患者数据或可识别个人信息而免于完整伦理审查（决定编号001，日期为2026年8月4日）。相关机构办公室也确认，使用和分析LLM生成的输出无需另行批准。分析依据LLM提供方的适用使用条款进行。发现：以安全为导向的语言在所有输出中占主导（93.3%），同时还有情感慰藉和怀旧。相比之下，明确的冲突、仪式具体性和强烈欲望优先化则始终缺席。即使在提示偏好偏离时，AI输出也倾向于中和张力，并将选择重新导向和谐、规避风险的叙事。结论：在所分析的全部输出中，生成式AI复制了一种狭隘的死亡道德脚本，它赋予安全和情感安宁以特权，同时边缘化冲突、仪式和具身欲望。我们并不声称临床医生目前依赖通用AI工具来决定患者的最后一餐；相反，这些发现说明，当偏好无法被直接确认时，此类工具可能如何充当非正式参考点，或在繁忙的临床环境中充当未经审视的“初稿”，以及当以这种方式被咨询时，此类工具默认采用的道德脚本。因此，护理伦理不仅必须审视AI推荐了什么，还必须审视此类建议所维持的照护与死亡的规范性愿景，并且必须将护士自身的道德能动性——而非AI输出——置于临终决策的中心。

英文摘要

BackgroundThe selection of an end-of-life meal for a person nearing death is often framed as a deeply personal and meaningful act. Distinct from the well-documented practice of last-meal requests before execution, food and meal choices during the terminal phase of illness are shaped by medical constraints, family dynamics, cultural expectations, and ethical judgements. Palliative-care nursing literature increasingly recognises food and mealtimes as sources of both meaning and distress for patients and families. As generative artificial intelligence (AI) systems are increasingly introduced into healthcare contexts, they may also generate recommendations in ethically sensitive areas such as end-of-life care. Such recommendations are not neutral; they may reflect normative assumptions about safety, comfort, and what constitutes a "good death."Research aimThis study examines how generative AI constructs and justifies end-of-life meal suggestions, and how these outputs reflect, reinforce, or silence particular ethical and cultural imaginaries of dying.Research designA qualitative, interpretive research design grounded in science and technology studies (STS) and critical AI-ethics perspectives on non-human actors as normative agents (see Theoretical framework, Section 4.1) was adopted, treating AI-generated outputs as cultural texts rather than clinical guidance. The analysis focused on both proposed meals and the justificatory language accompanying them.Participants and research contextNo human participants were involved. The dataset consisted of 75 AI-generated meal proposals produced by a single large language model. Fifteen systematically varied prompts were generated five times each, manipulating age, cultural background, religious orientation, family presence, swallowing difficulty, and preference deviation.Ethical considerationsThe authors' institutional research ethics committee confirmed that the study was exempt from full ethical review because it involved no human participants, patient data, or identifiable personal information (determination no. 001, dated 4 August 2026). The relevant institutional office also confirmed that no separate approval was required for the use and analysis of LLM-generated outputs. The analysis was conducted in accordance with the applicable terms of use of the LLM provider.FindingsSafety-oriented language dominated across outputs (93.3%), alongside emotional comfort and nostalgia. By contrast, explicit conflict, ritual specificity, and strong desire prioritisation were consistently absent. Even when preference deviation was prompted, AI outputs tended to neutralise tension and redirect choices towards harmonised, risk-averse narratives.ConclusionsAcross the outputs analysed, generative AI reproduced a narroaw moral script of dying that privileged safety and emotional tranquillity while marginalising conflict, ritual, and embodied desire. We do not claim that clinicians currently rely on generic AI tools to determine patients' final meals; rather, the findings speak to how such tools may function as an informal reference point when preferences cannot be directly confirmed, or as an unexamined "first draft" in busy clinical settings, and to the moral script such tools default to when consulted in this way. Nursing ethics must therefore interrogate not only what AI recommends, but the normative visions of care and death that such recommendations sustain, and must keep the nurse's own moral agency-rather than AI output-at the centre of end-of-life decision-making.