China's Cyberspace Administration has spent five years building a layered system for supervising algorithms rather than banning them outright. On July 15, 2026, that system grew a new layer. The Interim Measures for the Administration of AI Anthropomorphic Interaction Services — jointly issued April 10, 2026 by the 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 — took effect, imposing algorithm-filing, minor-protection, and anti-dependency obligations on any service that simulates a person's personality, thinking patterns, or communication style to sustain ongoing emotional interaction.
The measures don't stand alone. They sit on top of the Provisions on the Administration of Algorithm-Recommendation Services for Internet Information Services, effective March 1, 2022, which already requires algorithm providers with "public opinion properties or social mobilization capacity" to register with the CAC's algorithm-filing system within ten working days of launch (CAC, Jan. 4, 2022). Anthropomorphic-AI providers already filed under that regime must now update their filings to reflect human-like interaction features; new entrants face filing and safety-assessment obligations once they cross 1 million registered users or 100,000 monthly active users (Latham & Watkins). It's registry stacking, not registry replacement — a pattern that also runs through the 2023 deep-synthesis rules and the 2025 generative-AI content-labeling measures.
What the rules actually require
The substance is specific rather than aspirational. Providers must disclose, clearly and repeatedly, that a user is talking to AI rather than a person. Virtual "intimate partner" or "virtual relative" products are flatly barred for minors; any other anthropomorphic service offered to a user under 14 requires verified guardian consent. For adult users, providers must avoid designs that "excessively cater to users" or induce emotional dependency, and must push a pop-up reminder every two continuous hours of use warning that the content is AI-generated (Latham & Watkins). Human operators must be able to intervene when a user shows signs of self-harm risk. Training on user conversations requires opt-in consent.
The test of whether this bites came immediately. On the July 15 effective date, ByteDance's Doubao and Alibaba's Qwen both disabled their personalized AI-agent/companion features, notifying users the functions were offline for "product function adjustments" (SCMP). These are not marginal apps — Doubao and Qwen sit among China's most-used consumer AI products. That two of the country's largest platforms chose to pull functionality rather than adapt it in place within the compliance window is a strong signal that the filing and behavioral-design requirements are a real operating constraint, not a paperwork exercise regulators will quietly decline to enforce.
Steelmanning the rule
The strongest case for the measures doesn't require citing China's political system at all. Companion and emotional-support chatbots are a genuinely novel harm surface: unlike a recommendation feed, they're built to simulate relationship, and platforms have commercial incentive to make that simulation as sticky as possible. MIT Media Lab research has found some AI chatbots more addictive than social feeds precisely because they adapt to tell users what they want to hear. Character.AI and OpenAI have both faced lawsuits alleging their products contributed to a minor's suicide. Against that backdrop, a mandatory two-hour break reminder and a hard bar on virtual-partner products for children are proportionate, concrete, and — crucially — auditable in a way that vaguer duty-of-care standards are not. California's SB 243, which took effect in January 2026 with similar break-reminder and minor-protection provisions, and pending bills in New York and Washington state, show Beijing isn't alone in reaching this conclusion (Forbes).
Where it goes wrong
The problem isn't the anti-dependency substance — it's the delivery mechanism. By running these obligations through the same algorithm-filing architecture as the 2022 recommendation-algorithm registry, Beijing gives itself something California's law doesn't: a standing, granular map of exactly which companies operate emotionally interactive AI, how their models are configured, and what content policies they apply, all disclosed to a regulator with no independent judicial check on how that information gets used. A break-reminder mandate protects users regardless of who enforces it. A mandatory algorithm-configuration filing to a state cyberspace authority is a different kind of instrument — one that in China's system has previously doubled as a lever for content and viewpoint control, not just safety supervision. A five-agency joint order with mandatory registration thresholds, applied to a $30 billion-plus global product category still finding its safety norms, will chill exactly the smaller, non-dominant entrants the "inclusive and prudent" language claims to protect — while the largest platforms, with compliance teams built for the 2022 registry, absorb the cost and consolidate further. A narrower rule — mandatory disclosure, the break reminder, the under-14 bar — would have captured the genuine safety case without folding companion AI into a state visibility regime built for something else entirely.