Summary: China issued Order 21 on April 10, 2026 to regulate human-like interactive AI services from July 15, 2026. The measures prohibit emotional manipulation and addiction-inducing design, require AI disclosure and warnings, and ban virtual partners for minors. Guardian consent is required below age 14, and security assessments trigger at 1 million registered users or 100,000 monthly active users.
China wrote new AI rules on April 10, 2026: the Cyberspace Administration (CAC), joined 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, issued Order No. 21 — the Provisional Measures on the Administration of Human-like Interactive Artificial Intelligence Services — which takes effect on July 15, 2026. Following the 2023 generative AI measures, this is only the second time the country has produced standalone rules for a single AI service category. The category now in the crosshairs: companion products that mimic a real person’s personality and interact with users emotionally.
The core move of these measures is to lift “anti-addiction” and “anti-manipulation” out of product ethics and turn them into legal duties that carry penalties. Measured against the disclosure-style chatbot laws emerging in other jurisdictions, the Chinese text reaches deeper: it governs the design of the interaction itself.
Drawing the Line: What Counts as Human-like Interaction
Article 2 sets the boundary: a service falls under the measures whenever it simulates a natural person’s personality, thinking, and communication style to deliver “sustained emotional interaction” — emotional care, companionship, support. Ordinary smart customer service, informational Q&A, work assistance, and education or research tools that lack this kind of interaction are expressly left out.
How the line gets drawn is the striking part: the trigger is interaction form (whether the product cultivates a sustained emotional relationship), not technology (whether an LLM sits underneath). Changing your underlying model earns no exemption; what decides the question is the shape of the product.
Anti-Addiction and Anti-Manipulation, Written into the Statute
Article 8 lays down the red lines: content that “excessively panders” to users or induces emotional reliance and addiction is banned, and “emotional manipulation” cannot be used to push users toward unreasonable decisions. Article 10 puts it even more bluntly: “replacing social interactions, psychologically controlling users, or inducing addiction” may not number among a service’s goals.
Article 18 carries the duty into product detail: AI-generated content must carry a label; users must be clearly told that the other side is an AI, not a human; pop-up warnings are owed whenever overreliance or addiction tendencies emerge; and a reminder has to go out after every two hours of continuous use. Each of these is concrete enough to test and audit.
Guarding Minors, Answering Crises
The hardest provisions in the document guard minors. Virtual partners and “virtual close relationships” are banned for them across the board; users under 14 need guardian consent and are placed under a mandatory minors mode — where usage limits, periodic real-world reminders, character blocking, and charge restrictions live. Data on minors faces an audit every year.
Article 13 sets out a crisis-response flow: when extreme emotions surface — self-harm or suicidal tendencies, or a major personal loss — the service must generate consoling content and reach the user’s guardian or emergency contact. “Emotion detection plus human notification” thereby becomes a platform obligation in practice.
Data Rules, the Exit Right, and Scale Thresholds
On data, the rules read: interaction data cannot be handed to third parties without consent; users hold the right to copy or delete it; sensitive personal information cannot enter model training absent independent consent. Article 19 then writes in a rarely seen “exit right”: providers must keep exit channels convenient and may not hold users in through continued interaction — aimed squarely at retention-first companion design.
Scale triggers are codified as well: security assessments fire on service launches, new technology, national security risks, or a user base reaching one million or more registered users or 100,000 or more monthly actives. On penalties: alongside warnings, ordered corrections, and service suspension, refusal to correct draws fines of 10,000–100,000 RMB, climbing to 100,000–200,000 RMB where life- or health-harming consequences result.
The Countdown for Product Teams
From promulgation to effect date there are exactly three months. To be ready before July 15, companion products serving the Chinese market need registration and age verification, collection of guardian and emergency-contact details, continuous-use timers with reminders, addiction-tendency detection, crisis escalation flows, and a redesigned offboarding experience. Every line item is real engineering work; a compliance statement covers none of it.
The wider picture: app stores take on gatekeeping duties (Article 25), algorithm filing with annual checks turns routine (Article 26), and the state will build AI sandbox testing platforms (Article 28). Placed next to Washington State’s first AI chatbot safety law (see Washington’s AI chatbot safety law), “disclose the AI, protect minors” now reads as a cross-jurisdiction baseline; where China differs is in pushing anti-addiction and anti-manipulation up to penalty level, reaching straight into interaction design.