China's Interim Measures for the Administration of AI Anthropomorphic Interaction Services took effect July 15, 2026, and the response from the country's largest AI platforms was immediate: ByteDance's Doubao and Alibaba's Qwen disabled the persona and custom-agent features that let users build AI companions with a fixed personality, tone, and memory. Tencent's Yuanbao followed. These were not niche features — Doubao's user-created agents reportedly reached hundreds of millions of accounts before the shutdown, and users now have until October 15 to export their data before it becomes inaccessible.
The rules were issued jointly on April 10, 2026 by the Cyberspace Administration of China (CAC) alongside 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 — a five-agency alignment that signals this is treated as a cross-cutting governance priority, not a niche content rule. It is China's first regulation dedicated specifically to services that "simulate a person's personality traits, thought patterns, and communication styles to provide ongoing emotional interaction," distinguishing them from customer-service bots, tutoring tools, and research assistants, which are explicitly excluded from scope.
The Case for the Rule
The strongest argument for these Measures is not paternalism for its own sake — it is that anthropomorphic AI is a genuinely different product category from a search box or a customer-service chatbot, and the harms are already documented. Character.AI, the best-known Western analogue, has faced wrongful-death litigation in the United States over a minor's suicide linked to sustained chatbot interaction, and its own user base skews toward heavy, hours-long daily sessions among 18-24 year-olds. Xiaoice, China's largest AI companion service, has built a user base of roughly 660 million — evidence that emotionally engaging AI has already achieved mass-market penetration in China before any guardrails existed. A product engineered to sustain emotional attachment, deployed at that scale to a population that includes tens of millions of minors, is a reasonable target for sector-specific rules rather than reliance on general content law.
Measured against that problem, the Measures are more surgical than sweeping. Article 14 bars "virtual relatives, virtual companions," and other virtual intimate relationships for users under 18, and requires guardian consent specifically for users under 14 — a narrower, more calibrated line than an outright ban on the category. Article 18 requires a break reminder after two continuous hours of use, well short of a hard cutoff. Article 13's crisis-intervention requirement — generating de-escalating content and, in cases of imminent self-harm risk, contacting a guardian or emergency contact — targets the specific failure mode that has drawn the most public alarm about companion AI, rather than regulating tone or ideology. None of this bans the category for adults, and the rules explicitly encourage "orderly expansion" of anthropomorphic AI in eldercare and cultural applications.
Where the Rule Overshoots
The compliance response nonetheless shows the design has real costs. Doubao and Qwen did not build age-verification and guardian-consent flows and keep the feature running for adults — they shut it down entirely, for every user, citing routine "product function adjustments" rather than acknowledging the regulatory driver. That is the predictable outcome of Article 22's security-filing trigger, which requires a formal assessment once a service crosses 1 million registered users or 100,000 monthly active users — thresholds that any mainstream Chinese consumer app clears almost immediately after launch. Combined with algorithm review, ethics review, and content-management obligations layered on top, the compliance burden made "disable the feature" cheaper than "build the guardrails," even for firms with ByteDance's and Alibaba's engineering resources.
That is the proportionality question regulators elsewhere should study before copying this framework: a rule aimed at protecting minors ended up removing a legal, adult-facing product entirely, because the filing and review thresholds treated a companion feature the same as launching an entirely new AI service. The EU's approach under the AI Act's Article 5 manipulation provisions and the UK Online Safety Act's age-assurance duties both try to scope obligations to the actual risk tier rather than a blunt user-count trigger — worth emulating here. A narrower rule requiring age-gating and the specific minor protections in Article 14, without triggering the full Article 22 filing regime at a 100,000-MAU threshold, could plausibly have delivered the same minor-protection outcome without forcing companies to strip the feature from adults entirely.
The substance of what China legislated — mandatory AI disclosure, a two-hour reminder, crisis-intervention protocols, and a guardian-consent line for younger teens — is a defensible, evidence-grounded response to a real and growing harm. The lesson for other regulators drafting anthropomorphic-AI rules is in the implementation, not the intent: keep the compliance thresholds proportionate to actual risk, or watch companies delete the product instead of fixing it.