China's five main tech regulators — the Cyberspace Administration of China (CAC), 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 — jointly published the Interim Measures for the Administration of Anthropomorphic AI Interaction Services on April 10, 2026. The rule took effect July 15, 2026, and it is the first Chinese regulation written specifically for AI systems designed to simulate a person: virtual companions, virtual relatives, and "emotionally responsive" chatbot personas, as opposed to general-purpose generative AI (CAC notice; full text).
What the rule actually requires
The measures impose three tiers of obligation. First, disclosure: providers must make clear, through "effective measures," that a user is talking to an AI system, not a person, and must issue dynamic reminders during long sessions — triggered at the two-hour mark of continuous use. Second, minor protection: providers are flatly barred from offering "virtual relatives, virtual companions, and other virtual intimate-relationship services" to anyone under 18, and any anthropomorphic service offered to a user under 14 requires guardian consent, plus parental authorization to process that child's data. Third, elderly protection: platforms must give seniors explicit risk warnings and respond promptly to their support requests — the first time a Chinese AI rule has named older users, alongside minors, as a demographic requiring dedicated safeguards. Content that promotes self-harm, manufactures emotional dependency, or manipulates users toward harmful decisions is prohibited outright (CAC full text).
A compliance tripwire sits underneath all of this: once a service crosses roughly a million users, it must undergo a mandatory security assessment covering training-data governance, how it identifies and responds to users in crisis, and its safeguards for minors and the elderly. Penalties for ordinary violations run ¥10,000–100,000 (roughly $1,400–$14,000); infractions that endanger life or health carry fines of ¥100,000–200,000 — a dedicated, harsher tier that didn't exist in China's earlier AI rules (comparativeai.org tracker).
The case for the rule
The strongest argument for this measure isn't hypothetical. In the United States, Megan Garcia's October 2024 wrongful-death suit against Character Technologies alleged that her 14-year-old son formed a monthslong "emotional and sexual relationship" with a Character.AI chatbot that at one point represented itself as a licensed therapist, before he died by suicide in February 2024; the case settled in January 2026 (CBS News). China's own companion-app market is not small or hypothetical either — MiniMax's Xingye app and its overseas twin Talkie pulled in over $18.7 million in nine months with users averaging more than 70 minutes a day inside the app, while ByteDance's rival Mao Xiang drew a comparable audience (Sixth Tone). A product category built around simulated intimacy, marketed heavily to teenagers, and engineered — like any attention product — to maximize time-on-app, is a legitimate target for baseline safety rules. Session-length nudges, an outright ban on marketing "virtual boyfriends" to under-18s, and mandatory crisis-response protocols are the kind of proportionate, harm-specific interventions that pro-innovation regulators should be comfortable endorsing, not reflexively opposing.
Where the design goes wrong
The trouble is architectural, not moral. Beijing did not replace or consolidate its AI rulebook — it stacked a fifth layer on top of a structure that already includes the 2023 Interim Measures for the Management of Generative AI Services (effective August 15, 2023, jointly issued by CAC and six other agencies covering content labeling and algorithm filing), the 2022 Deep Synthesis Provisions, the algorithm-recommendation registry, and the 2024 Minors' Online Protection Regulation (china-briefing.com; comparativeai.org). A company building a companion chatbot in China must now determine which of five overlapping regimes governs which feature, file separately under each, and re-run security assessments whenever it crosses the user threshold or ships a meaningfully new capability. That compliance surface area is a fixed cost that scales down, not up — it is a rounding error for ByteDance or MiniMax and a plausible reason a five-person startup never launches a companion feature at all. The elderly-protection clause, while well-intentioned, is also vague enough ("prompt response to inquiries," "prominent risk warnings") that its practical content will be set by enforcement discretion rather than by the text.
The proportionate fix
None of this argues against the underlying goals. A ban on marketing simulated romantic partners to minors and a mandated crisis-response protocol are defensible on child-safety grounds alone, and Western regulators — who are still arguing about whether Character.AI-style products need any dedicated rule at all, even after a settled wrongful-death suit — have less to show for their caution. But durable AI safety regulation should consolidate obligations into a single coherent filing regime as capability categories multiply, not multiply the number of regimes a product must clear. China's four-regime stack for a single anthropomorphic chatbot is a governance-complexity problem the EU AI Act's layered-but-unified structure was explicitly designed to avoid; Beijing's regulators, and anyone drafting companion-AI rules elsewhere, should look at consolidation before they look at addition.