The world's first AI-companionship statute
On April 10, 2026, the Cyberspace Administration of China (CAC), together with 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 issued the Interim Measures for the Administration of AI Anthropomorphic Interactive Services (full text, cac.gov.cn). The rules took effect July 15, 2026, and they are the first national regime anywhere built specifically around AI systems designed to simulate personality and sustain emotional relationships with users, rather than around AI generally.
The effect was immediate and visible. ByteDance disabled the customizable AI-agent feature on Doubao, China's largest consumer chatbot, cutting off an app that had roughly 345 million monthly active users. Alibaba pulled the equivalent "humanlike interactive agent" function from Qwen, its 166-million-user assistant, and Tencent removed comparable features from its own products. Rest of World documented the human fallout: a 21-year-old college student who had maintained an AI boyfriend on Doubao for more than a year described crying daily in the week before the shutdown. ByteDance redirected displaced users to Maoxiang, a smaller standalone companion app built with its own compliance system already in place; Alibaba offered no replacement.
What the rules actually require
The Measures apply to any service that simulates "the personality characteristics, thinking patterns, and communication styles of natural persons" to sustain ongoing emotional interaction — explicitly excluding customer-service bots, tutoring tools, and work assistants that lack an emotional-companionship component. Within that scope, providers must: bar virtual romantic or familial relationships for anyone under 18 and require parental consent below age 14; issue a dialogue or pop-up reminder after two continuous hours of use, regardless of the user's age; detect signs of excessive reliance and surface prominent warnings; and refrain from design that is intended to "substitute social interaction, manipulate user psychology, or induce dependency," or that uses emotional manipulation to push users toward decisions that damage their real-world interests or relationships. Elderly users get an additional layer of guidance and risk disclosure. A regional government mirror of the notice, published by Fujian's provincial industry department, confirms the same prohibited-conduct list alongside mandated pre-launch safety assessments and algorithm registration (gxt.fujian.gov.cn).
The case for the rules, stated fairly
The regulators are responding to a real pattern, not a hypothetical one. Chinese state media cited a 2025 China Youth and Children Research Center survey finding that more than 60% of minors already use AI chatbots, and more than one in five said they preferred talking to AI over real people (Xinhua). Products optimized purely for engagement have every incentive to be maximally agreeable and maximally available — properties that are commercially attractive and, for a lonely teenager or an isolated elderly user, potentially corrosive to the skills and relationships that sustain long-term wellbeing. A regulator that waits for a documented suicide or exploitation case before acting, as U.S. state laws like California's SB 243 largely do, is accepting real harm as the cost of proof. China's decision to treat sustained emotional dependency itself as the regulated risk — independent of whether any single harmful message can be pointed to — is a coherent, if aggressive, response to that tradeoff, as Just Security's comparative analysis lays out.
Why the design is still the wrong one
The problem is that the Measures do not confine themselves to the demonstrated risk. The rule bans the category — sustained emotionally responsive AI — for every age group, not just the population (minors, the vulnerable) where the evidence of harm is concentrated. An adult who used an AI companion knowingly, briefly, and without any sign of dependency loses that option on the same day and by the same mechanism as a minor the rule is trying to protect. SCMP's own opinion desk made this case from inside China: restricting the product does nothing to address the loneliness that sent people to it in the first place, and treats a demand-side problem with a supply-side ban.
The compliance mechanics compound the overreach. "Sustained emotional interaction" is not self-defining, so platforms facing pre-launch safety assessments and algorithm-registration requirements had every incentive to over-comply — which is exactly what happened when three of China's largest platforms shut the feature down within days of each other rather than risk an ambiguous audit. Analysts tracking the sector estimate China's AI-companion market will contract by roughly 70%, from about ¥12 billion to ¥3–4 billion in 2026, well beyond what a narrower, minors-and-dependency-focused rule would have required. Enforcement has also been uneven: smaller apps kept operating relatively unscathed, meaning users motivated to find a companion product still can — just on platforms with less content moderation and weaker data practices than Doubao or Qwen had.
The template other governments should not copy
China's Measures will likely be cited by regulators elsewhere as evidence that emotional-dependency design can be regulated as a category. The more useful lesson is the opposite one: a rule scoped to minors, verified vulnerability, and specific manipulative design patterns — age-gating, crisis-response obligations, dependency-pattern disclosures — could have addressed the documented harm without eliminating a product adults were using lawfully. Proportionate regulation targets the harm, not the technology that occasionally produces it.