A new product category, regulated from the top down
On July 15, 2026, China's Interim Measures for the Administration of Anthropomorphic Artificial Intelligence Interaction Services took effect, roughly three months after the Cyberspace Administration of China (CAC) published the final text jointly 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 on April 10, 2026 (CAC). The rules single out a specific product behavior — AI that simulates personality and sustains emotional engagement over time — rather than regulating chatbots generally, carving out customer-service and productivity tools from scope.
The compliance response was immediate and visible. ByteDance's Doubao, China's largest consumer AI app, and Alibaba's Qwen stripped out user-built persona and companion features ahead of the deadline, with ByteDance redirecting users to a separate, more restricted companion app (Just Security; Caixin Global).
The case for the rule
The strongest argument for these measures isn't hypothetical. Regulators worldwide, including in the US and EU, have grappled with reports of AI companion apps fostering unhealthy dependency in teenagers, and in isolated cases, contact with chatbots preceding self-harm. China's approach treats emotional dependency as a design problem rather than a content problem: providers must build in two-hour continuous-use reminders, cannot architect systems to "excessively cater to" users in ways that displace human relationships, and must flag clearly, at every session, that the user is talking to software, not a person (CAC Q&A). For a product category expanding faster than most parents or regulators can track, mandating these guardrails at the platform-design level — rather than waiting for harm and litigating after the fact — is a defensible, arguably more effective, model than the US's reactive, suicide-focused liability approach.
The minor-specific provisions are the measures' clearest achievement. Anthropomorphic services cannot offer "virtual kinship" or "virtual companion" roles to minors at all, and any anthropomorphic interaction with a user under 14 requires verified parental consent, with mandatory child-safe modes, reality reminders, and guardian-visible spending and role controls (CAC Q&A). These are close to bright-line rules, not vague standards, and they map onto a specific, well-documented harm: children forming exclusive attachments to simulated "family members" or romantic partners inside apps with no adult oversight.
Where the measures overreach
But the rules bundle that legitimate child-safety architecture with content restrictions that have nothing to do with emotional dependency. Providers are separately barred from generating content that "endangers national security," "incites subversion of state power," or reflects "ideological deviation" (CAC) — categories with no fixed legal definition, adjudicated at the discretion of the same agency enforcing the minor-protection rules. A companion chatbot that drifts into a politically sensitive conversation with an adult user faces the same compliance exposure as one that fails to intervene in a mental-health crisis. Conflating the two doesn't make either rule better enforced; it gives platforms every incentive to over-censor ordinary conversation to avoid an ambiguous political violation, since the cost of guessing wrong on "ideological deviation" is far higher than the cost of an overcautious chatbot.
The crisis-intervention mandate raises a related concern. Providers must "identify safety risks" when a user shows signs of extreme emotional distress and contact emergency contacts or guardians (CAC Q&A) — a genuinely protective idea for, say, a suicidal teenager. But it also requires platforms to build systematic detection and reporting of users' emotional states, with no public detail on how that data is stored, who else can access it, or how long it's retained. Analysts have flagged this as creating the legal infrastructure for state visibility into individual psychological risk markers, packaged as a safety feature (geopolitechs). A narrower rule — mandatory crisis-response protocols with strict data-minimization and no state-reporting channel — would have captured the safety benefit without the surveillance externality.
A better model exists
The pre-launch security-assessment requirement, triggered once a service crosses roughly 1 million registered or 100,000 monthly active users, also stacks compliance cost specifically on the mid-sized firms trying to compete with Doubao and Qwen, entrenching the incumbents that can absorb it. A risk-tiered rule that scaled review intensity with actual harm indicators — reported self-harm incidents, minor-user share — rather than raw user count would better serve both safety and competition.
China got the core insight right: emotional-dependency risk in AI companions is a design problem that benefits from ex ante rules, and minors deserve hard limits, not just disclosure. But by folding that insight into an omnibus content-control regime with vague national-security triggers and unbounded crisis-data collection, Beijing has built a rule that will be remembered as much for what it lets the state see as for what it protects children from.