Paternalistic AI: the case of aged care

The paper examines whether AI used in aged care can unjustifiably constrain older adults’ autonomy under the guise of safety and well-being. Against the backdrop of rapidly aging populations and anticipated care shortages, AI is promoted to extend and enhance care. The authors propose that ageism—stereotypes depicting older adults as frail, vulnerable, and incompetent—can shape the design and deployment of AI, leading to technological paternalism that restricts freedom without consent for the person’s supposed good. They outline their argument: identifying how age bias enters AI through design scripts and data, explaining how this fosters paternalism, defining technological paternalism, and evaluating whether paternalistic AI toward older adults is justified.

The authors synthesize emerging work on AI ageism, defining it as AI practices that discriminate against or neglect older adults’ interests. Building on scholarship about AI bias (e.g., Buolamwini and Gebru; Noble), they highlight that ageism is pervasive yet understudied in AI. They review: (1) age scripts in design, where technologies for older adults are often built by younger developers with stereotyped assumptions, focusing on health and safety while overlooking leisure and autonomy; older adults are frequently excluded or minimally involved in design processes with biases toward healthier, tech-savvy participants. Case studies (e.g., AIMS remote monitoring) show scripts that limit older adults’ ability to resist or reappropriate technologies, though some users negotiate privacy and adapt systems. Market dynamics further bias designs toward customers’ priorities (family, providers) over end-users’ autonomy. (2) Biased data: older adults are underrepresented in datasets due to exclusionary data collection and gray digital divides; clinical research often omits older cohorts or lacks age-disaggregated data, undermining generalizability and performance for older users. These dynamics risk misinterpreting individual variation as aberrant and reinforce ageist stereotypes.

Conceptual and normative analysis. The authors delineate and apply philosophical accounts of paternalism (Dworkin; Mill; soft vs hard paternalism) to AI in aged care. They introduce and analyze technological paternalism, assessing whether AI systems satisfy three conditions of paternalism: limiting liberty, doing so purportedly in the user’s best interests, and without consent. They draw on illustrative cases from the literature (e.g., smart homes, monitoring, fall detection, medication management) to ground the analysis. They then evaluate justifications for paternalistic interventions via competence-based (soft) and welfare-maximizing (hard) frameworks, incorporating arguments about life-year opportunities and heterogeneity among older adults to assess when, if ever, paternalistic AI is warranted.

• AI systems for aged care can satisfy the conditions of paternalism: they can restrict older adults’ actions and choices (e.g., through surveillance, automated alerts, access controls, rigid medication schedules); are designed and used to promote health and safety as the conception of the user’s good; and often operate without explicit, informed consent or effective opt-out/override control by users.
• Ageism enters AI via age scripts and biased data. Design practices frequently exclude older adults or treat them as a homogeneous, vulnerable group, encoding stereotyped assumptions that prioritize safety over autonomy. Data underrepresentation and lack of age disaggregation impair model performance for older users and misclassify normal individual variation as risky.
• Market/customer dynamics prioritize the preferences of payers (families, care organizations) for safety and monitoring, over end-users’ autonomy and preferences, further embedding paternalistic features.
• Soft paternalism is generally unjustified for older adults solely on the basis of age: many older people are competent decision-makers with rich life experience and may reasonably value autonomy, privacy, and meaningful risk-taking. Competence varies and must be assessed case-by-case.
• Hard paternalism weakens with age when assessed by life-year opportunities: the putative benefits of restricting freedom to extend life diminish with age, and paternalistic interventions may reduce meaning and pleasure more than they add safety for many competent older adults.
• Therefore, blanket or universal imposition of paternalistic AI in aged care is not justified. Paternalistic AI may be appropriate for subgroups lacking decision-making capacity, but imposing it broadly on competent older adults is unjust.

The analysis shows that AI in aged care can entrench ageist assumptions and produce technological paternalism that constrains older adults’ autonomy. While AI promises efficiency and safety in the face of workforce shortages, systems often privilege health and risk avoidance over user-defined values. This misalignment arises from design scripts, exclusion of older adults from design, biased datasets, and market incentives privileging customers’ risk perceptions. The findings support a shift toward participatory, context-sensitive design that includes older adults early and continuously, ensures meaningful control and consent, and recognizes heterogeneity in capacities and preferences. Ethical assessment should balance safety with autonomy, privacy, and quality of life, and avoid treating advanced age as a proxy for incompetence. Where paternalistic AI is considered, justifications must be individualized, transparent, and proportionate, with user override and opt-out mechanisms to mitigate autonomy infringements.

AI for aged care frequently embeds ageist scripts and data biases that yield technological paternalism. The authors argue that, as a general rule, imposing paternalistic AI to promote older adults’ overall good is unjustified: many older adults are competent and may reasonably prioritize autonomy and meaningful risk. Paternalistic interventions may be warranted for those lacking decision-making capacity, but not as blanket policy. They call for participatory design with older adults, careful consideration of value trade-offs (safety vs autonomy), case-by-case assessment based on specific health conditions and decision-making abilities, and rigorous attention to inclusive, age-disaggregated data and user control. Future research should empirically investigate real-world impacts of AI systems on older adults’ autonomy and well-being and develop governance that prevents ageist and paternalistic defaults.

The paper is a conceptual, normative analysis without new empirical data. The authors acknowledge their diagnosis is not universal: some AI systems are inclusively designed and adaptable, and older adults may creatively appropriate technologies. Use contexts vary, and not all deployments are coercive or paternalistic. Findings should be applied case-by-case, considering heterogeneity among older adults and specific system designs.