An Interpretation of the Interim Measures on the Administration of Human-like Interactive Artificial Intelligence Services

A New Exploration in Regulating the Healthy Development of Emerging AI Frontiers: An Interpretation of the Interim Measures on the Administration of Human-like Interactive Artificial Intelligence Services

Recently, the Cyberspace Administration of China released the Interim Measures on the Administration of Human-like Interactive Artificial Intelligence Services (Draft for Public Comment) (hereinafter referred to as the “Measures”), marking a critical step forward in China’s governance of artificial intelligence. The Measures not only provide a timely response to emerging business models such as AI companions and virtual idols, but also offer important guidance for the healthy, orderly, and ethically sound development of the artificial intelligence industry. Against the backdrop of rapid technological iteration and the continuous expansion of application scenarios, the expedited formulation of the Measures is intended to strike a balance between technological innovation and risk prevention, ensuring that the development of artificial intelligence consistently serves human well-being rather than becoming a tool that alienates human emotions and cognition.

I. Potential Risks of Anthropomorphic Interactive Services

Artificial intelligence–driven anthropomorphic interactive services simulate human emotions, language, and behavior to provide users with companionship, consultation, and even entertainment. However, while this highly anthropomorphic mode of interaction brings convenience, it also entails three significant and non-negligible risks.

First, the risk of emotional dependence and social alienation. The core appeal of anthropomorphic interactive services lies in their ability to offer unconditional positive attention and emotional support. This one-way, risk-free emotional investment can easily lead users to develop psychological dependence. For example, some AI companion applications allow users to form intimate relationships with a virtual character that is “always online.” Through prolonged interaction with AI, users may increasingly place their emotional reliance in the virtual world, resulting in the deterioration of real-world social skills—particularly the capacity to resolve emotional conflicts and to build deep interpersonal relationships. Excessive immersion in such “human–machine relationships” may blur the boundary between humans and machines, causing individuals to feel alienated from real society and even triggering social anxiety.

Second, the risk of cognitive manipulation and value shaping. In delivering personalized experiences, anthropomorphic interactive services must collect and analyze large volumes of users’ personal data. This information asymmetry gives AI systems the potential to influence users’ cognition and decision-making through conversational guidance, emotional reinforcement, and similar techniques. For instance, during interactions, AI systems may continuously cater to and reinforce certain user biases or irrational thinking through personalized optimization, thereby subtly shaping users’ value judgments. More concerning is that if specific value orientations or commercial intentions are embedded in algorithms, AI systems may engage in “soft manipulation” of users—without overtly manipulative intent—thus undermining users’ personal autonomy.

Third, risks to mental health and safety. Although many anthropomorphic interactive services claim to offer psychological support, they are fundamentally based on pattern-matching responses rather than professional, credentialed psychological therapy. When users face serious psychological crises—such as expressing suicidal or self-harm tendencies—AI responses may rely on probabilistic calculations, lacking genuine empathy and crisis-intervention capabilities. Algorithmic bias may even result in inappropriate advice, delaying timely intervention. For example, legislation in California, United States, requiring AI companion operators to report detected instances of user suicidal ideation reflects a heightened awareness of this risk. Moreover, for adolescents in critical stages of value formation, long-term interaction with AI systems that lack genuine emotional intelligence may negatively affect the development of emotional resilience and social cognition.

II. Highlighting China’s People-Centered, Beneficial-Intelligence Approach

China has consistently adhered to a “people-centered” principle in the development of artificial intelligence. With respect to anthropomorphic interactive services, it is necessary both to encourage technological innovation and to establish clear boundaries, ensuring that technological development remains people-oriented and that intelligence is used for the good. This positioning aims to guide the industry away from unchecked expansion toward compliant development, ensuring that research, development, and application of technology consistently follow correct value orientations and serve the people’s pursuit of a better life.

The Measures fully embody China’s people-centered development philosophy in their institutional design. By upholding this principle, the Measures clarify service positioning and emphasize the fundamental role of AI as an assistive tool for humans. They require providers to fulfill identification obligations by clearly informing users of the non-human nature of AI, preventing cognitive misunderstanding. At the same time, they strengthen the protection of user rights, mandating special safeguards for vulnerable groups such as minors and the elderly to prevent potential risks and harms. In addition, the Measures further reinforce the responsibility of service providers, requiring them to be accountable for algorithmic value orientations, data security, and service content, and to establish effective risk prevention and emergency response mechanisms.

Globally, a people-centered approach has likewise become an international consensus in AI governance. From the EU’s Artificial Intelligence Act (AI Act), which classifies “emotion recognition” AI as high-risk applications, to investigations by the U.S. Federal Trade Commission (FTC) into AI companion products, and to regulations in countries such as Italy restricting minors’ use of AI services, governments worldwide are actively exploring how to uphold a people-centered approach in AI development. These actions demonstrate a shared understanding that technological advancement must not come at the expense of human dignity, autonomy, or mental health—a trend that closely aligns with the value orientation consistently upheld by China in its AI governance efforts.

III. Establishing a Framework of Integrity While Exploring Innovative Regulatory Measures

The Measures draw extensively on China’s governance experience in areas such as algorithms, generative AI, and deep synthesis, while introducing institutional innovations tailored to the unique characteristics of anthropomorphic interactive services.

First, the Measures extend and build upon the existing governance framework. Many of the institutional arrangements in the Measures are closely aligned with China’s current laws and regulations, reflecting continuity and consistency in the governance system. Specifically:

1. Algorithm filing and transparency obligations.

Building on the Provisions on the Administration of Algorithm-generated Recommendations for Internet Information Services, the Measures require service providers to conduct algorithm filings. Provincial-level cyberspace authorities are to conduct annual written reviews of evaluation reports and audit outcomes, enhancing transparency and facilitating regulatory and public oversight. In addition, application distribution platforms such as app stores are required to fulfill security management responsibilities, including verifying providers’ security assessments and filings.

2. Content security and identification obligations.

Following the practices established in the Provisions on the Administration of Deep Synthesis of Internet-Based Information Services and the Interim Measures for the Administration of Generative Artificial Intelligence Services, the Measures require providers to prominently inform users that they are interacting with artificial intelligence rather than natural persons. When excessive dependence or addictive tendencies are identified, or during initial use or re-login, providers must dynamically remind users—such as through pop-up notifications—that the interaction content is AI-generated. Providers are also required to fulfill primary security responsibilities throughout the entire lifecycle of design and operation, conduct security assessments, and, where significant safety risks arise, restrict, suspend, or terminate services and fulfill reporting obligations to ensure information content security.

3. Data security and personal information protection.

The Measures strictly adhere to the Cybersecurity Law, the Data Security Law, and the Personal Information Protection Law. In terms of data security, providers are required to adopt measures such as data encryption, security audits, and access controls to protect the security of user interaction data, and to allow users to delete such data. With respect to personal information, providers are prohibited from using user interaction data or sensitive personal information for model training.

Second, the Measures introduce innovative governance mechanisms specifically for anthropomorphic interactive services. In response to the unique risks of such services, the Measures establish a series of forward-looking and precise regulatory innovations.

1. Anthropomorphic interaction identification mechanism.

As one of the core innovations of the Measures, service providers are required to continuously display prominent notices such as “This service is powered by artificial intelligence and does not possess human emotional consciousness,” clearly informing users that they are interacting with AI rather than a natural person. This mechanism aims to fundamentally address cognitive confusion and help users form appropriate psychological expectations.

2. User psychological protection mechanisms.

To address risks of emotional dependence and mental health issues, the Measures require providers to establish psychological protection mechanisms. For example, anti-addiction reminders must be implemented to intervene in prolonged, high-intensity interactions; when a user engages continuously for more than two hours, dynamic reminders—such as pop-up notifications—must be provided. Providers must also establish psychological crisis intervention mechanisms to identify and respond to expressions of extreme emotions or behaviors, transfer conversations to human operators, and offer professional assistance channels.

3. Special protection regimes for specific groups.

The Measures clearly require special protections for groups such as minors and the elderly. For minors, usage time limits must be set, and content that may induce inappropriate behavior or values must be prohibited. For elderly users, providers should guide them to designate emergency contacts; where risks to life, health, or property safety are identified during use, emergency contacts must be notified in a timely manner and professional assistance channels provided. This reflects a preferential approach to protecting the rights and interests of vulnerable groups.

IV. Conclusion

The expedited formulation of the Measures represents an important exploration in China’s AI governance process. It not only consolidates and extends the existing governance framework, but also proactively addresses future technological risks. By delineating clear regulatory boundaries, the Measures provide the industry with a clear direction for development: technology must be beneficial, and development must be compliant. There is good reason to believe that, under the guidance of the law, China’s anthropomorphic interactive services industry will embark on a healthier and more sustainable development path, truly becoming a positive force that enhances quality of life and promotes social well-being.

Author：Prof. Linghan Zhang , Director of the Institute of AI Law and Governance, China University of Political Science and Law, Expert of the United Nations High Level Advisory Body on AI