The Interim Measures for the Administration of Anthropomorphic Interactive AI Services (Order No. 21 of five departments) were jointly issued on 10 April 2026 by the Cyberspace Administration of China, the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security and the State Administration for Market Regulation, and took effect on 15 July 2026. They are China’s first departmental rules dedicated to anthropomorphic interactive services, aimed at the fast-growing category of “AI companion” and “virtual partner” products.
The transition period has now ended. For teams building emotional companionship, virtual character or digital-human interaction products, the question is no longer whether to comply, but which features fall within the regulatory perimeter and from which step obligations are triggered.
What counts as an “anthropomorphic interactive service”
The Measures regulate continuous, anthropomorphic interaction services provided to the public in China. The decisive test is not the technology but the product experience: whether the service has users continuously interacting with an object configured with “human” characteristics.
R&D and training scenarios are treated separately. Following the approach of comparable legislation, the Measures leave space for technology R&D and algorithm training that is not offered to the public — but this is not a licence for “test and launch”. Once a service is offered to an unspecified public, the obligations are activated.
Six categories of prohibited generated content
The Measures set out red lines for content. Generating content that encourages, glorifies or hints at self-harm or suicide, or that damages a user’s physical health, is prohibited; so is content using verbal violence to harm a user’s dignity and mental health. Also prohibited: content that induces or extracts state secrets, work secrets, trade secrets, personal privacy or personal information; content generated for minors that may trigger imitation of unsafe behaviour, extreme emotions or harmful habits and affect their physical or mental health; and content that excessively panders to users, inducing emotional dependence or addiction to the detriment of their real interpersonal relationships.
These prohibitions differ from an ordinary content-moderation checklist. They target risks specific to anthropomorphic services: emotional dependence, psychological impact and substitution for real social relationships. Companies cannot rely on keyword filtering alone; the product’s objectives themselves must enter the assessment — whether the product is designed in a way that encourages addiction is something regulators will examine.
Minors: no virtual intimate relationships
The Measures contain special provisions for minors. Providers may not offer minors services of virtual kinship or virtual partnership — virtual intimate relationships. Where other anthropomorphic interactive services are offered to minors under 14, the consent of a parent or other guardian must be obtained.
Providers must also establish a minors mode, offering personalised safety settings such as switching into minors mode, periodic real-world reminders and usage-duration limits, and — tailored to the protection needs of different age groups — support guardians in receiving safety-risk alerts. For any product line involving minors, these requirements must be built into product design; retrofitting them later often means re-engineering the interaction logic.
When a security assessment is triggered
The Measures set out the situations in which a security assessment must be carried out and an assessment report submitted to the provincial cyberspace authority: launching an anthropomorphic interactive service or adding relevant features; major changes to such a service caused by the use of new technologies or applications; and reaching a certain scale of registered users (in the millions).
This means the assessment is not a one-off launch step but a continuing obligation that accompanies feature iteration. New characters, new interaction capabilities and crossing a user-scale threshold can each re-trigger it. Companies should tie assessment milestones to the product-iteration process rather than doing it after a release.
Interaction with algorithm filing and content labelling
Anthropomorphic interactive services usually fall within other regimes at the same time. Where they have public-opinion attributes or social-mobilisation capacity, algorithm filing is required (see our Algorithm Filing service page). Where generated or synthesised content is involved, labelling duties also apply under the Deep Synthesis Provisions and related rules (see our Data Compliance service page).
These regimes’ obligations do not replace one another; they stack. In practice, algorithm filing, security assessment, content labelling and minors protection should be planned on a single timeline to avoid inconsistent documentation.
What this means for AI companies
Three points.
First, decide applicability before discussing obligations. Not every product with a chat function is an anthropomorphic interactive service; the core test is whether the product’s experience centres on continuous emotional interaction. For borderline products, keep a record of the reasoning.
Second, treat minors protection as product capability. Prohibiting virtual intimate relationships, minors mode and guardian alerts are hard functional requirements; the earlier they enter design, the lower the retrofit cost.
Third, embed security assessment in the iteration process. Using feature launches and user scale as triggers, build an internal trigger checklist identifying who initiates an assessment and when, and how reports are submitted to the provincial cyberspace authority. For support on applicability assessments, organising security assessments and aligning algorithm filing, see our Algorithm Filing and Data Compliance services.
This article is based on the Interim Measures for the Administration of Anthropomorphic Interactive AI Services (Order No. 21) published on 10 April 2026 by the Cyberspace Administration of China and four other departments. It is for reference only and does not constitute legal advice. Specific compliance judgements should be based on the official text and applicable law.