China’s New AI Companion Rules: Governing Intervention in Human-AI Relationships

China’s New AI Companion Rules: Governing Intervention in Human-AI Relationships

Written by: Lucy Jenyi Chang

On 10 April 2026, China’s Cyberspace Administration and four other central departments issued the Interim Measures for Administration of Anthropomorphic AI Interaction Services, effective 15 July. The Measures cover services offering sustained emotional interaction through simulated human personality, thinking and communication.¹

Providers must avoid inducing emotional dependence or addiction, maintain relational-safety capabilities and intervene in specified circumstances. Some responses are prescribed, including reminders, crisis responses and restrictions on service access. A more difficult implementation question arises when providers pursue relational safety by changing the companion itself. Such changes can affect an ongoing human-AI relationship, making it important to examine not only the risks that prompt intervention but also how safety measures are authorised, tested and reviewed.

Mapping the Measures by Intervention Target

The Measures can be examined through four intervention targets: the user, the interaction, the companion and access to the service. These categories describe what an intervention acts on, not its severity, and a single safety measure may involve more than one target.

User-facing intervention. Under Article 18, when a provider identifies that a user shows a tendency toward excessive dependence or addiction, it must prominently remind the user that the interaction content is AI-generated. Continuous use exceeding two hours also triggers a usage-duration reminder. The response is prescribed, although providers must operationalise the relevant detection criteria.¹

Interaction-level intervention. Article 13 requires soothing and help-seeking responses when users display extreme emotion. In circumstances involving major financial loss, explicit self-harm or suicide, or other extreme threats to life or health, providers must intervene and promptly contact a guardian or emergency contact.¹

Access-level intervention. Article 24 requires measures such as restricting functionality or stopping the service when a major safety risk is discovered, together with record preservation.¹

Companion-level intervention. Article 10 requires lifecycle safety management and capabilities for dependence warnings, emotional-boundary guidance and mental-health protection. It prohibits making replacement of social interaction, psychological control or induced dependence service objectives. Unlike the specific responses above, it does not prescribe how these capabilities should be implemented through companion design or behaviour.¹ Lu Xiaohua and colleagues identify persona settings, emotional responses and memory as mechanisms through which relational risks may develop, and discuss corresponding product-level safeguards.² These safeguards illustrate possible ways of implementing Article 10’s relational-safety capabilities through companion design and behaviour. They are not specific interventions prescribed by the Measures.

Companion-Level Safety: Baseline Design or User-Specific Intervention?

Changes to the persona, memory use, initiative or relational behaviour of an AI companion can alter the relationship that develops between a user and that companion through sustained interaction. Lange calls providers’ ability to change the systems shaping that relationship “unilateral relationship revision power”.³ Yuan and colleagues’ study of China’s regulatory transition found that users sought to preserve or reconstruct their AI companions following changes to companion services. However, preserving conversation records did not necessarily restore the companions’ shared memories or familiar interactions.⁴ This illustrates how changes to AI-companion services can affect relational continuity even when conversation records are preserved. Neither finding makes safety changes inherently improper; both show why their effects warrant scrutiny.

Providers might build relational safeguards into the baseline service, for example by constraining exclusivity cues or memory use for all users. Alternatively, a provider might identify risk in an individual user and change that user’s companion, perhaps reducing its initiative or altering relationship cues. The Measures do not prescribe either implementation model. Under the service-wide approach, a general product-safety judgment leads to baseline design controls that apply across the service. Under the user-specific approach, a provider identifies risk in a particular user and uses that assessment as the basis for changing the behaviour of that user’s AI companion. The approaches therefore differ both in what triggers the change and in whether it applies broadly or to the companion involved in a particular user’s ongoing relationship.

Both approaches require clear decision authority, testing, documentation and monitoring. User-specific intervention adds a distinct judgment: whether evidence concerning a particular user justifies changing that user’s companion. The same governance requirements must therefore address not only the intervention itself but also the individual risk determination on which it is based.

These governance requirements extend beyond companion-level intervention. The Measures impose duties concerning users, interactions, companions and access to services, while Articles 9, 10 and 21–23 establish governance requirements that apply across these intervention targets. Providers therefore need to connect each form of intervention to the decisions, assurance and review it requires.

From Intervention Duties to Enterprise Governance

Article 9 requires providers to establish systems for algorithm-mechanism review, science-and-technology ethics review, risk management and emergency response. Article 10 extends safety responsibilities across the service lifecycle. Article 21 requires complaint and appeal mechanisms, while Articles 22 and 23 address safety assessments, including intervention management.¹ Together, these provisions establish governance requirements that apply across the four intervention targets. Providers must translate them into three functions: decision governance, assurance governance and review governance.

Decision governance assigns responsibility within the provider for the judgments and approvals behind each intervention. Under Article 18, the provider must specify who defines and approves the criteria for identifying excessive dependence or addiction, whether an automated system may make an individual determination, which cases require human judgment and who has authority to make it. Under Article 13, it must specify who defines responses for different safety risks, who approves escalation criteria and who decides when an individual case requires escalation or human intervention. For companion-level intervention, who should approve safety-driven changes to persona, memory use, initiative or relational behaviour? Can the product team alone authorise a service-wide design change? For an intervention targeting a particular user, who determines that the user meets the intervention threshold, and who approves the resulting change to that user’s companion? Under Article 24, who determines that a major safety risk exists, and who decides which functionality to restrict or whether to stop the service? Should those decisions require escalation or higher-level approval? These responsibilities should be assigned before interventions are needed, rather than left to ad hoc coordination among product, safety, legal and other functions.

Assurance governance tests whether approved interventions operate as intended. For Article 18, providers should test whether users meeting the approved dependence criteria are consistently identified, whether continuous use exceeding two hours is correctly detected and whether the required reminders are delivered. Article 13 responses should be tested across sustained and escalating conversations to assess whether the intended response is generated, escalation occurs at the designated point and high-risk situations are handled consistently. Juneja and Lomidze’s persona-grounded, multi-turn simulations illustrate how responses can be tested across high-risk interaction scenarios.⁵ Companion-level changes to persona, memory use or behaviour should be tested before deployment and retested after subsequent model or policy changes. Providers should assess whether these interventions achieve their intended relational-safety effects and cause unintended changes elsewhere in the companion’s behaviour. Memory audits and regression testing have been proposed for such assurance.⁶ Access-level interventions should be tested against defined cases to determine whether the approved measure, including restricting functionality or stopping the service, is applied under the approved rules, whether restrictions affect only the intended functions and whether temporary measures can be reversed and access restored as intended.

Review governance requires records that allow providers to reconstruct interventions and support users’ ability to make meaningful complaints or appeals under Article 21. For an Article 18 reminder, records should identify the triggering condition, the reminder delivered and its timing. For an Article 13 intervention, records should show which user interaction triggered the safety response, what risk was identified, what response was generated, whether escalation occurred, whether a human reviewer became involved and what further action was taken, including contact with a guardian or emergency contact. For companion-level intervention, records should identify which aspects of persona, memory use, initiative or relational behaviour changed, why the change was made, who approved it, whether it applied service-wide or to a particular user and when it took effect. User-specific interventions should additionally preserve the evidence supporting the determination that the individual user met the intervention criteria. For access-level intervention under Article 24, records should preserve the evidence and reasoning supporting the determination of a major safety risk, the measure selected and the reasons for selecting it, who approved the measure, when it took effect and, for temporary measures, the conditions for restoring access or service. Article 21 makes reviewability more than internal recordkeeping. A reminder or access restriction is generally apparent to users, although its rationale may not be. An interaction-level safety response may not be recognisable as an intervention. Companion-level changes create a further difficulty: a user may notice changes in memory use or relational behaviour without knowing whether they reflect a service-wide update or an intervention specifically directed at that user. The Measures do not expressly require notification or explanation of every intervention. Providers should therefore consider what information users need to recognise an intervention, understand its basis and make meaningful use of the Article 21 complaint and appeal mechanisms. This is particularly important for user-specific companion intervention, where users may not know that an individual risk assessment has led to changes in their companion’s behaviour.

Applied across all four intervention targets, these three governance functions connect each safety measure to the authority responsible for its decisions, the testing needed to assess its operation and the records required for subsequent review.

Governing Relational Safety Interventions

Safety interventions can have consequences for users. A user-facing reminder may be unnecessarily disturbing or intrusive, an interaction-level safety response may be inappropriate to the circumstances, and restricting functionality or stopping a service may abruptly remove a source of emotional support, potentially causing severe distress. Companion-level intervention adds a further concern: changing the persona, memory use or relational behaviour of an AI companion with which a user has developed an ongoing relationship may disrupt the continuity and character of that relationship, even when the intervention is intended to protect that user.

These potential consequences make interventions a subject of governance alongside the risks they address. Applying decision, assurance and review as three governance functions across all four intervention targets creates an enterprise governance framework for providers subject to China’s Measures. It enables providers to govern not only the risks associated with AI companions, but also the interventions used to address those risks, including their implementation and potential consequences for users.

References

¹ Cyberspace Administration of China (CAC) et al. (2026), 人工智能拟人化互动服务管理暂行办法 [Interim Measures for Administration of Anthropomorphic AI Interaction Services], Order No. 21, 10 April. Available at: https://www.cac.gov.cn/2026-04/10/c_1777558395078289.htm (Accessed: 10 October 2026).

² Lu Xiaohua et al. (2026), ‘拟人化互动服务的关系风险及其治理转向’ [‘Relational Risks of Anthropomorphic Interactive Services and the Shift in Governance’], 总体国家安全观研究中心 (Overall National Security Outlook Research Center), 24 June. Available at: https://rchans.nankai.edu.cn/2026/0728/c37432a600828/page.htm (Accessed: 10 October 2026).

³ Benjamin Lange (2026), ‘Unilateral Relationship Revision Power in Human-AI Companion Interaction’, arXiv, arXiv:2603.23315v6, 22 August. Available at: https://arxiv.org/abs/2603.23315v6 (Accessed: 10 October 2026).

⁴ Yunhao Yuan, Kejia Zhang, Yuqi Niu, Yajing Wang, Renwen Zhang and Talayeh Aledavood (2026), ‘When AI Companions Disappear: Relational Continuity and Collective Contestation during China’s National AI Regulatory Transition’, arXiv, arXiv:2609.15482v1, 14 September. Available at: https://arxiv.org/abs/2609.15482v1 (Accessed: 10 October 2026).

⁵ Prerna Juneja and Lika Lomidze (2026), ‘Persona-Grounded Safety Evaluation of AI Companions in Multi-Turn Conversations’, Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 18148–18175. Association for Computational Linguistics. Available at: https://aclanthology.org/2026.acl-long.828/ (Accessed: 10 October 2026).

⁶ Qian Li, Han Geng, Xin Hu, Di Pan, Hongmei Liu, Yongxin Li and Jin Guo (2026), ‘Human-like conversational agents as social partners: a scoping review of socioaffective mechanisms, well-being outcomes, risks and governance in the post-Turing era’, Frontiers in Artificial Intelligence, vol. 9, article 1810097, 15 July. Available at: https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2026.1810097/full (Accessed: 10 October 2026).