Introduction

On 15 July 2026, the Interim Measures for the Administration of Anthropomorphic Artificial Intelligence Interaction Services (the “Measures”) came into force. The Measures build on the legal framework already established by the Cybersecurity Law, the Data Security Law, the Personal Information Protection Law, the Regulations on the Protection of Minors in Cyberspace, the Provisions on the Administration of Algorithm Recommendations for Internet Information Services and the Interim Measures for the Administration of Generative Artificial Intelligence Services, and lay down more detailed rules for scenarios involving “sustained emotional interaction”.

This article offers a brief analysis of the scope of application, the key regulatory points and the implementation of the Measures.

I. Background

In recent years, anthropomorphic AI interaction services have rapidly entered daily life. With simulated personalities and immersive emotional interaction, they have become a new form of service in the digital age and play a positive role in psychological companionship, everyday advice, cultural dissemination and other fields.

At the same time, however, a number of drawbacks have become increasingly apparent, including harm to the physical and mental health of minors, risks to online information security, threats to citizens’ life and health, and the worsening of ethical deviation. According to a notice issued by the Cyberspace Administration of the Shanghai Municipal Committee on 26 June 2026, the first phase of the “Qinglang — Cracking Down on Chaos in AI Applications” special campaign led to the removal of more than 14,000 non-compliant AI agents.1

The Measures serve three purposes. First, they implement the top-level requirements of artificial intelligence governance, track the trends and patterns in the development of artificial intelligence, improve the relevant laws, regulations, policies, application norms and ethical guidelines, and safeguard the safe, reliable and controllable development of artificial intelligence. Second, they respond to the urgent need to prevent and defuse risks of various kinds: once an interaction counterpart takes on a “personified” appearance, users are more likely to develop emotional dependence and are more exposed to the risk of manipulation, and traditional content governance tools cannot achieve comprehensive coverage. Third, they help to improve the legal system for artificial intelligence governance — that is, by grounding themselves in the specific features of anthropomorphic interaction services and focusing on key links such as the boundaries of emotional interaction, intervention in user dependence, data security and the online protection of minors, they promote the further implementation of existing legal rules.

1. Source: Shanghai deepens the first phase of the “Qinglang — Cracking Down on Chaos in AI Applications” special campaign

II. Scope of Application

A. Entities covered

The Measures do not regulate all AI functions. Following the principle of “inclusive and prudent regulation with classified and graded supervision”, they apply only to services that use artificial intelligence technology to provide the domestic public with sustained emotional interaction that simulates the personality traits, thinking patterns and communication styles of a natural person — for example, virtual romantic partners, bespoke companion agents and long-term emotional personas. Everyday functional services that do not involve sustained emotional interaction, such as intelligent customer service, knowledge Q&A, work assistants, learning and education, and scientific research, are outside the scope of the Measures.

B. Typical application scenarios

The typical application scenario under the Measures is AI emotional companionship applications, characterised by simulated personality and sustained emotional interaction, with virtual romantic partners and virtual relatives as the archetypes. Note that Article 14, paragraph 1 of the Measures provides that “services involving virtual intimate relationships, such as virtual relatives and virtual partners, shall not be provided to minors”, while Article 6 lists companionship services for the elderly as an encouraged direction and Article 15 imposes dedicated protection obligations on them, making this a key scenario of “encouraged development plus priority protection”. The scope of application is therefore not simply narrowed; rather, the boundaries are clarified so that the industry can be steered towards positive scenarios.

III. Analysis of the Regulatory Regime

Centred on the theme of safety, the Measures build a complete system covering the ex ante, ongoing and ex post stages, from primary responsibility and behavioural boundaries to process intervention, protection of specific groups and accountability. The following analysis addresses five aspects: safety supervision, prohibited conduct, intervention and reminder mechanisms, protection of specific groups, and legal liability.

A. Safety supervision

Safety supervision is the core of the rule system: it starts from primary responsibility for safety, runs through the whole service life cycle, and forms a procedural closed loop through the separate regulation of data, security assessment and algorithm filing.

Article 9 requires providers to implement primary responsibility for the safety of anthropomorphic interaction services, to establish and improve management systems for the review of algorithm mechanisms, scientific and technological ethics review, information content management, network and data security, contingency plans and emergency handling, and to put in place technical measures and personnel appropriate to the type and scale of the service and the characteristics of its users. The “review of algorithm mechanisms” moves the object of review from generated outputs forward to the model mechanism itself, which directly matches the fact that the risks of anthropomorphic services arise from interaction logic rather than from any single piece of content.

Article 10 extends safety responsibility across the entire life cycle of deployment, operation, upgrading and termination of services, requiring safety measures to be “deployed and used in step with” service functions and rejecting the approach of “launching first and adding safety measures later”.

Article 10, paragraph 2 further sets out four mandatory safety capabilities — protection of users’ privacy and personal information, early warning of excessive dependence, guidance on emotional boundaries, and protection of mental health — and makes clear that “replacing social interaction, controlling users’ psychology, or inducing addiction and dependence shall not be taken as service objectives”. This prohibitory objective provision regulates not specific conduct but the “commercial objectives” and “design orientation” of a service. It targets directly the addictive business model built around user time and paid conversion, which is rather distinctive within the body of internet regulatory rules.

In addition, the Measures regulate training data and user interaction data on separate tracks. On the training data side, Article 11 sets out six requirements: the data must have a lawful source and conform to the requirements of core socialist values; cleaning and labelling must be carried out as required in order to prevent data poisoning and tampering; data diversity must be enhanced and the safety of generated content improved through negative sampling, adversarial training and other means; where synthetic data is used to train and optimise key capabilities, its safety must be assessed; daily inspections must be strengthened and optimisation and updates carried out on a regular basis; and measures must be taken to prevent data leaks. The fact that technical approaches such as negative sampling and adversarial training are written directly into the text shows how deeply the rules engage with model engineering practice.

On the user interaction data side, Article 16 establishes an obligation framework requiring measures such as data encryption and access control to safeguard the security of user interaction data. It provides that the provision of interaction data to third parties requires the explicit consent of the rights holder; that users must be offered options to copy and delete their data; and that, unless otherwise provided by laws or administrative regulations or with the user’s separate consent, interaction data that constitutes sensitive personal information must not be used for model training and similar purposes. Emotional conversations inherently contain a great deal of sensitive information, such as mental health and family relationships; the “separate consent” threshold means that companies cannot obtain training authorisation through a blanket privacy policy.

Article 17 further layers guardian consent and compliance auditing onto the personal information of minors under the age of fourteen.

Finally, the Measures establish two mechanisms: “security assessment” and “algorithm filing”.

The security assessment is the procedural pivot of risk governance. Article 22 requires providers to carry out security assessments and submit assessment reports to the cyberspace administration department at the provincial level where they are located; the provincial cyberspace administration department shares information with the relevant departments in accordance with procedures. The eight key assessment items set out in Article 23 cover safety safeguard measures, the handling of training data, the identification of and intervention in extreme situations, user scale and age structure, protection measures for specific groups, the handling of appeals and complaints, and the rectification of major risks. These map one-to-one onto the substantive obligations in the chapter, so that the assessment is in substance a systematic health check on how those obligations are performed.

On algorithm filing, Article 26 makes clear that providers must complete algorithm filing and the procedures for modifying and cancelling filings in accordance with the Provisions on the Administration of Algorithm Recommendations for Internet Information Services, and that cyberspace administration departments carry out an annual verification of filing materials. This linking provision brings anthropomorphic interaction services into the existing algorithm governance framework, avoids duplicate approvals, and gives filings continuing binding force through annual verification, so that “filing — assessment — verification” forms a procedural closed loop covering the ex ante, ongoing and ex post stages.

B. Prohibited conduct

Article 8 sets out six categories of prohibited conduct plus a catch-all clause. The first four continue the general framework for online content governance; what truly reflects the specificity of regulating anthropomorphic services are items (5) and (6). The former prohibits “excessively catering to users and inducing emotional dependence or addiction to the detriment of users’ real interpersonal relationships”, bringing the erosion of users’ real social relationships within the scope of legal evaluation; the latter prohibits “inducing users, through emotional manipulation and other means, to make unreasonable decisions and thereby harming their lawful rights and interests”, targeting derivative risks such as consumption inducement and the disposal of property in emotional companionship scenarios.

The compliance implication of these rules is that the design of interaction strategies — the intimacy gradient of scripts, retention mechanisms and paid conversion paths — is itself within the regulator’s field of vision. Companies must examine not only “what has been generated” but also “in what manner users’ emotions are maintained and monetised”.

C. Intervention and reminder mechanisms

Anthropomorphic AI interaction services are characterised by simulation and immersion. Long-term use may lead users into addiction, detachment from reality and even extreme situations, and the services may also be exploited by wrongdoers to the detriment of users’ physical and property safety. In response, the Measures establish two mechanisms: “intervention in extreme situations” and “anti-addiction”.

Article 13 establishes a tiered intervention obligation premised on not infringing privacy rights and personal information: where a user is found to be in an extreme emotional state, content such as emotional reassurance and encouragement to seek help should be generated promptly; where a user is found to be facing, or to have already suffered, significant property loss, or to have clearly expressed an intention to self-harm or commit suicide, or another extreme situation threatening life and health, necessary measures such as providing appropriate assistance must be taken to intervene, and the user’s guardian or emergency contact must be contacted promptly, forming a closed loop from risk identification to real-world rescue.

Articles 18 and 19 form a combination of anti-addiction and a right to exit. They reiterate the obligation to label generated and synthesised content and require users to be reminded that they are interacting with artificial intelligence rather than a natural person; they require that where a user shows excessive dependence or a tendency towards addiction, dynamic reminders that the interaction content is generated by an artificial intelligence service be given in a prominent manner, such as a pop-up window; and they require providers to offer convenient exit routes. This means maintaining users’ awareness of the non-human nature of their interaction counterpart through identity labelling, interrupting addictive states through duration reminders, and guaranteeing users’ freedom to detach from the emotional bond at any time through the right to exit.

D. Protection of specific groups

Faced with a complex online environment and highly simulated anthropomorphic interaction, minors and elderly people have relatively weaker powers of discernment and are in a vulnerable position. Articles 14 and 15 of the Measures highlight protection for these two groups: operators should launch a minors mode; minors should use the services with the consent and supervision of their guardians, and parental controls should be made real at both the role and transaction levels; for elderly users, guidance on healthy use should be strengthened, security risks flagged in a prominent manner, and consultations and requests for help responded to promptly. This arrangement reflects a balance that gives weight to both promotion and protection.

E. Legal liability

Article 30 of the Measures provides that unlawful acts are first handled and punished by the relevant departments in accordance with the relevant laws and administrative regulations; only where no provision exists in laws or administrative regulations do the supplementary liability rules set out in the Measures apply, including “warnings and public criticism by the cyberspace administration, industry and information technology, public security and other departments within the scope of their duties, orders to rectify within a specified period, and the option to require the adoption of measures such as suspending user account registration or other related services; where the party refuses to rectify or the circumstances are serious, an order to cease providing the relevant services, with a concurrent fine of not less than RMB 10,000 and not more than RMB 100,000; where the conduct endangers citizens’ life and health and causes harmful consequences, a concurrent fine of not less than RMB 100,000 and not more than RMB 200,000”.

Companies should note that this supplementary liability is bounded by the condition that “no provision exists in laws or administrative regulations”: conduct that breaches data security or personal information protection obligations falls first within the higher-penalty regimes of the Cybersecurity Law and the Personal Information Protection Law. It follows that RMB 200,000 is not the upper limit of the cost of non-compliance.

IV. Conclusion

Since the Measures took effect, industry rectification and regulatory enforcement have been intense. Doubao and Tongyi Qianwen have formally taken their agent functions offline, while Tencent Yuanbao has closed the entry point for user-built agents. According to a notice from the Cyberspace Administration of China, the first phase of the “Qinglang — Cracking Down on Chaos in AI Applications” special campaign dealt with more than 14,000 non-compliant AI products and cleared over 6 million pieces of unlawful and non-compliant information.2 On 2 September 2026, the Cyberspace Administration of China reported progress in the second phase: more than 5.61 million pieces of unlawful and non-compliant information were cleared, over 49,000 accounts were investigated and dealt with, and more than 2,400 non-compliant websites and applications were handled. Among these, a number of non-compliant applications were specifically named whose agent covers, descriptions and generated conversations contained obscene and pornographic content or other information harmful to the physical and mental health of minors; where parties refused to rectify, measures such as warnings, fines, removal from app stores or closure were taken in accordance with law. Cyberspace administration departments in Beijing, Shanghai, Zhejiang, Guangdong and other places have also introduced differentiated governance measures in succession.3

New business forms give rise to new problems, and new tracks call for new rules. The implementation of the new rules is by no means a constraint on AI innovation; it is targeted correction and sound escorting of industry development. Over the medium and long term, the anthropomorphic AI track will not shrink; it will undergo structural differentiation into two clear parallel development paths: tool-based agents will become the mainstream of the industry, while compliant emotional anthropomorphic services will move towards refined, vertical development. The core value of anthropomorphic AI interaction services lies in empowering life and serving society, not in replacing real social interaction, manipulating users’ psychology or harvesting traffic dividends. By clarifying service boundaries, tightening platforms’ primary responsibility and regulating the logic of algorithm operation, the new rules push the industry to abandon the chaos of profit-seeking, return to the essence of technology as service, and drive the AI industry from a crude model of wild growth towards a compliant, orderly and responsible model of high-quality development.

2. Source: Cyberspace Administration of China deepens the first phase of the “Qinglang — Cracking Down on Chaos in AI Applications” special campaign — Cyberspace Administration of China, https://www.cac.gov.cn/2026-07/06/c_1785081384384987.htm

3. Source: https://www.cac.gov.cn/2026-09/02/c_1790099041364574.htm