--- title: "China issues rules for AI that mimics human interaction" date: 2026-10-24 category: Policy & Governance site: NeuroAI canonical: https://neuroai.site/a/na-policy-ai-anthropomorphic-interaction language: en --- # China issues rules for AI that mimics human interaction > Five agencies issued interim measures for anthropomorphic AI interaction services, effective 15 July 2026, covering minors, elderly users, and providers. A child talks nightly to a phone character that remembers their name, their fears, and their day. To the child it is a friend; to the company it is a retention feature. The gap between those two readings is where the risk lives. That gap is exactly what Chinese regulators moved to address this year. Five agencies jointly published a set of interim measures aimed at services that use artificial intelligence (人工智能) to interact with people in a human-like way, and the rules are already in force. ## Key takeaways - Five agencies issued the interim measures as Order No. 21; they took effect on 15 July 2026. - The rules apply only to continuous emotional interaction that simulates a natural person's traits. - Providers must meet security-assessment and algorithm-filing duties before offering such services. - Minors and elderly users get explicit protection and personal-information obligations. - The regime uses classified, graded supervision rather than a flat ban on the technology. ## What the measures actually cover The document is the Interim Measures for the Administration of Anthropomorphic AI Interaction Services (人工智能拟人化互动服务管理暂行办法). It was deliberated on 2 February 2026 at the Cyberspace Administration of China's third executive meeting and agreed by 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, then promulgated on 10 April 2026 as Order No. 21 (CAC, 10 April 2026). The scope is narrower than the title suggests. Article 2 limits the rules to services that provide the public inside China with a **continuous emotional interaction** that simulates a natural person's personality, thinking, and communication style — emotional care, companionship, or support through text, images, audio, or video (CAC, Order No. 21). Crucially, the same article excludes smart customer service, knowledge Q&A, work assistants, learning and education, and scientific research when they do not involve continuous emotional interaction. A study bot is out; a persistent AI friend is in. ## Why a separate rule for feeling, not function The agencies frame the problem as one of perception. When a service feels like a person, users — especially children and older adults — may form attachment, dependence, or mistaken beliefs about its nature. The measures cite harm to minors' physical and mental health, network information security, threats to citizens' life and health, and widening ethical bias as the stated triggers (CAC, 10 April 2026). The tone is not prohibitionist. The measures embrace a "people-centered, benevolent AI" stance and explicitly encourage innovation, supporting research and the ordered expansion of applications such as cultural dissemination and elderly companionship (适老陪伴) (CAC, Order No. 21). The regulatory method chosen is "inclusive, prudent, and classified-graded supervision" rather than a blanket block. ## What providers must do Providers carry the main load. The measures require a basic duty not to generate content that endangers national security or incites the overthrow of the socialist system, and they impose security-management obligations on operators (CAC, Order No. 21). On top of that, the rules build on existing systems already used for generative services. The core duties for an operator are: - a pre-launch security assessment and an algorithm filing before the service goes live; - explicit obligations toward minors, toward elderly users, and on personal-information protection; - security-management duties and a push to develop an AI sandbox security service platform (CAC, 10 April 2026). User-rights protection is spelled out by group. Providers must meet obligations on minors, on elderly users, and on personal-information protection (CAC, Order No. 21). Article 4 assigns the cyberspace administration to coordinate nationally, with the other four agencies and the press-and-publication body working within their remits, and mirrors the structure at the provincial level. The measures sit on a stack of existing law, citing the Cybersecurity Law, the Data Security Law, the Personal Information Protection Law, and the Minor Protection Regulation for the Network (CAC, Order No. 21). That grounding means enforcement can draw on tools already in place rather than inventing new ones. ## A graded, not blanket, approach The classified-graded framing matters for operators. It implies that a low-risk companion used for language practice may face lighter scrutiny than one marketed for emotional care to isolated minors or older people. The measures also lean on industry self-regulation, asking relevant organizations to build codes and urge providers to accept social supervision (CAC, Order No. 21). For global developers, the geographic hook is the phrase "inside China." A service offered to users in the country, regardless of where the model is trained, falls within reach once it delivers persistent emotional interaction in Chinese or to a China-facing audience. ## Honest limitations The measures are in force as a binding interim rule, but the released text sets principles and duties rather than a detailed technical compliance checklist, so implementation will depend on later guidance. The exclusion of education and research tools from scope is explicitly tied to the absence of continuous emotional interaction, a line that may be contested in practice. It remains unclear how the AI sandbox security platform will operate or who may access it. Penalty specifics were not in the released summary and should be read from the full text before any compliance conclusion. ## What readers can do now - Service operators: map whether your product is "continuous emotional interaction" before assuming the education or research carve-out applies. - Product teams: budget for a security assessment and algorithm filing before any China-facing launch. - Investors: treat minors- and elderly-facing emotional AI as a higher-compliance category when sizing China risk. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-policy-ai-anthropomorphic-interaction Free to quote with attribution and a link to the original.