A robotic baby harp seal named PARO has been comforting elderly patients in care homes for two decades, reducing anxiety and agitation in people with dementia across more than thirty countries. Yet according to a new comparative study published in AI & Society, the same unassuming device becomes three entirely different regulatory objects depending on where it is deployed. In Japan, it is a relational technology whose risks are woven into the social fabric of care. In China, it is a population-scale behavioural intervention demanding preventive standardisation. In Hong Kong, it is a data-processing system governed through legal adjudication. The finding carries a provocative implication: there is no universal patient, and the ethical categories regulators rely on—harm, vulnerability, accountability—are not properties of the technology at all, but products of governance itself.
The study, authored by Rosa E. Martín-Peña of Leibniz University Hannover and Zhengmin Li of the University of Copenhagen, challenges the dominant assumption in medical AI ethics that frameworks built on autonomy, individual rights and data protection can simply be applied worldwide. Global reviews of AI ethics documents have identified a stable “consensus package” of principles drafted mainly in Europe and North America, now embedded in guidance from the World Health Organization and the OECD. The authors argue that this dominance makes it easy to treat assumptions about personhood and harm as universal, when in fact what counts as harm, care, or proper human–machine interaction depends on local moral traditions, institutional arrangements and regulatory histories.
To make the comparison rigorous, the researchers developed a triadic analytical framework examining three dimensions in each jurisdiction: moral anthropologies, or how persons and harms are conceptualised; models of legitimacy, or who is authorised to define risk; and regulatory instruments, spanning the continuum from soft law to hard law. Japan, China and Hong Kong were selected as a most-different-systems design, representing soft-law cooperative governance, centralised hard-law architecture, and a common-law rights-based framework respectively. The payoff of this comparison is a fourth, unexpected axis: each jurisdiction embodies a distinct temporal logic of harm prevention—iterative in Japan, preventive in China, and reactive in Hong Kong.
Japan constructs the patient as a relational node within family and community rather than a bounded individual. This anthropology draws on Confucian role ethics, the philosopher Watsuji Tetsurō’s concept of ningen as a dialectical unity of individual and social existence, and the value of wa, or harmony, which appears explicitly in Japanese AI governance documents such as the Social Principles of Human-Centric AI. Legitimacy is technico-consensual, deriving from consensus among experts, industry and government in a pattern consistent with the nemawashi tradition of building agreement before formal decisions. Accountability is diffuse, spread across developers, providers, families and the state, and coordinated primarily through non-binding guidelines rather than enforceable mandates.
The temporal consequence is striking. Japan’s governance rationality trusts norms to emerge through use and professional dialogue, stipulating no harm-prevention requirements in advance of deployment. PARO itself, despite receiving US Food and Drug Administration certification as a Class II biofeedback medical device in 2009, operates in Japan without equivalent classification. Companion robots fall outside the Software as a Medical Device framework, so adverse relational outcomes—dependency formation, displacement of family contact, caregiver burden—are not systematically collected under any regulatory channel. Ethnographic research in Japanese care facilities documented one resident who became so attached to PARO that she refused to eat or sleep without it, a case of dependency that existing soft-law instruments do not directly address. The structural risk, the authors note, is that commercial deployment outpaces the norm-building process on which the model depends.
China takes the opposite approach, framing emotional dependency as a population health risk requiring preventive state intervention. The patient is constructed as a protected subject, and legitimacy derives from the Party-state’s role as guarantor of collective welfare and social stability. Hard-law instruments are layered densely: the Personal Information Protection Law and Data Security Law of 2021, National Medical Products Administration guidelines for medical AI, and sector-specific measures on algorithmic recommendation, deep synthesis and generative AI. The centrepiece is the Interim Measures for the Administration of Anthropomorphic AI Interactive Services, finalised in April 2026 and effective from 15 July 2026, among the most explicit attempts globally to regulate emotional AI directly. Key provisions include two-hour usage reminders to prevent dependency, restrictions on training models with interaction data, mandatory emergency contact mechanisms for elderly users, and a ban on AI systems simulating family relationships—a provision the authors read as consistent with Confucian role ethics, even though official commentary frames it in people-centred vocabulary.
This preventive architecture carries its own trade-offs. A system like PARO entering the Chinese market would face exhaustive ex-ante review across multiple regulatory regimes, yet a patient who developed problematic dependency would have no institutionalised avenue for contestation. The authors describe the result as mis-regulation rather than over-regulation: exhaustive procedural coverage of identifiable technical risks alongside a structural absence of frameworks for relational and emotional harms. Patients are protected yet not necessarily empowered, with administrative risk absorption substituting for individual redress. Enforcement practice illustrates the logic: in April 2026, the Cyberspace Administration of China took action against three prominent generative AI platforms for failing to implement mandatory content labelling, reflecting broader regulatory concern with emotional dependency and manipulation.
Hong Kong, by contrast, constructs the patient as a rights-holder with autonomy over health decisions and personal data, governed through a juridico-procedural rationality that has become constitutionally unsettled since 2020. The Personal Data (Privacy) Ordinance provides the binding baseline, imposing consent, purpose limitation and data security obligations, while a suite of non-binding instruments—the Privacy Commissioner’s ethical AI guidance and the 2024 Model Personal Data Protection Framework—translates principles into operational expectations. A 2025 compliance check of sixty organisations found no contravention of the Ordinance, and roughly 63 percent of organisations collecting personal data through AI referenced the Model Framework—but adoption remained voluntary. The temporal logic is reactive: violations are adjudicated only after they occur. A robot found to collect biometric data beyond its stated purpose would face regulatory action; a robot that fostered unhealthy emotional dependency or displaced family contact in a cognitively vulnerable resident would not, because those harms fall structurally outside the privacy perimeter.
The comparison exposes blind spots that no single jurisdiction escapes. Psychological dependency, relational disruption and the erosion of human empathy fall outside traditional regulatory categories in all three cases, differing only in which gaps are structural and which are contingent. The problem extends beyond Asia: the EU AI Act prohibits manipulative systems and those exploiting vulnerabilities, yet the European Commission’s own implementation guidelines suggest that a system detecting users’ moods and recommending emotionally aligned content may be permissible if it causes no significant harm—leaving the boundary between therapeutic care and prohibited manipulation indeterminate. The authors also warn of a potential “Beijing Effect”: because developers seeking access to the Chinese market must comply with usage limits and the family-simulation ban, these requirements are likely to reshape global product design, producing de facto convergence on China’s construction of emotional harm without any democratic deliberation about whose ontology of emotion should govern.
The study’s prescription is not harmonisation but epistemological humility: recognition that different traditions construct emotion, personhood and harm differently, and that no single construction has privileged access to truth. The authors propose concrete mechanisms, including disclosure of the governance rationality shaping product design through ISO/IEC standardisation processes, the UN Global Digital Compact’s review mechanisms, and bilateral regulatory dialogues such as the EU–Japan Digital Partnership. Such transparency would not resolve incompatibilities, but it would make them legible. As the authors conclude, governance without ontological awareness risks regulating technologies while misunderstanding the kinds of patients they are said to protect—and international frameworks will otherwise continue to govern a universal patient who does not exist.
Subject of Research: Comparative governance of emotional AI in healthcare across Japan, China and Hong Kong
Article Title: No universal patient: temporal logics of harm prevention in emotional AI healthcare across Japan, China, and Hong Kong
Article References: Martín-Peña, R. E., & Li, Z. (2026). No universal patient: temporal logics of harm prevention in emotional AI healthcare across Japan, China, and Hong Kong. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03355-5
Image Credits: AI Generated
DOI: 10.1007/s00146-026-03355-5
Keywords: emotional AI, companion robots, PARO, AI governance, healthcare ethics, Japan, China, Hong Kong, harm prevention, temporal logics, Confucian ethics, data protection
Cite Scienmag News
APA MLA Chicago
Denise Maddox. (September 25, 2026). One Robot, Three Rules: How Japan, China and Hong Kong See AI Harm Differently. Scienmag. https://scienmag.com/one-robot-three-rules-how-japan-china-and-hong-kong-see-ai-harm-differently/
Denise Maddox. “One Robot, Three Rules: How Japan, China and Hong Kong See AI Harm Differently.” Scienmag, 25 September 2026, https://scienmag.com/one-robot-three-rules-how-japan-china-and-hong-kong-see-ai-harm-differently/. Accessed 25 September 2026.
Denise Maddox. “One Robot, Three Rules: How Japan, China and Hong Kong See AI Harm Differently.” Scienmag. September 25, 2026. https://scienmag.com/one-robot-three-rules-how-japan-china-and-hong-kong-see-ai-harm-differently/
Copy citation Download RIS
Tags: AI governanceAI regulation differencesChinaChina’s population-scale AI interventionscompanion robotscomparative study of AI ethicsConfucian ethicscross-cultural AI policy analysiscultural perspectives on AI harmdata protectionemotional AIethical implications of robotic caregivingglobal AI regulation challengesharm preventionhealthcare ethicsHong KongHong Kong data governance in AIimpact of governance on AI categorizationinfluence of social context on AI deploymentJapanJapan’s relational AI approachPAROrobotic therapy in elderly caretemporal logics
