A registry built for recommendation engines now polices digital companions
China's Interim Measures for the Administration of Anthropomorphic AI Interactive Services (人工智能拟人化互动服务管理暂行办法), approved by the Cyberspace Administration of China (CAC) on February 2, 2026 and jointly issued with the National Development and Reform Commission, Ministry of Industry and Information Technology, Ministry of Public Security, and State Administration for Market Regulation on April 10, 2026, took effect July 15. It is China's first regulation written specifically for AI that simulates a personality — virtual companions, personality-mimicking chatbots, emotional "digital relatives," and role-playing game characters built for sustained emotional engagement, as distinct from customer-service bots or workplace assistants that aren't trying to be your friend.
The mechanism is old, even if the target is new. Article 26 requires providers to complete algorithm filing, plus any changes or deregistration, under the 2021 Provisions on Algorithm Recommendation of Internet Information Services — the rule that created China's mandatory algorithm registry in the first place. Rather than stand up parallel machinery for companion AI, Beijing folded a fast-growing, emotionally sensitive product category into infrastructure built five years ago to police news feeds and short-video recommendation engines.
What actually changed on July 15
Three provisions carry the weight. Article 22 requires a formal security assessment — covering training-data handling, crisis-response capability, and user demographics, among eight areas — whenever a provider launches a new anthropomorphic feature, makes a major technical change, or crosses 1 million registered users or 100,000 monthly active users. Article 18 requires a usage-time nudge, a pop-up or in-conversation reminder, after every two hours of continuous interaction, plus dynamic prompts if a user shows signs of dependence. Providers are barred from offering virtual-companion or virtual-relative services to users under 14 without verified guardian consent, with spending and time limits layered on top. Sensitive personal data drawn from a user's interactions cannot be used to train models without separate, specific consent.
Steelman the rule before dismissing it. China's digital-human and AI-companion market was valued at roughly $600 million in 2024, up 85% year-on-year, and products built to remember a user, mirror their tone, and never disagree are, by design, optimized for engagement in ways ordinary chatbots aren't.
Regulators citing "excessive catering" that induces "emotional dependence or addiction" — one of seven practices the measures explicitly prohibit — are naming a real product-design failure mode, not a hypothetical one. A rule forcing crisis-detection pathways and a break-reminder on a 14-year-old two hours into a conversation with a synthetic companion is not obviously disproportionate.
The tell is in how companies responded
What's notable is that ByteDance and Alibaba didn't retrofit their products — they killed the feature. Doubao took its customized humanlike-agent function offline by July 15, giving users a read-only window to view (not bulk-export) their agent configurations and chat histories until October 15, after which the data becomes unrecoverable. Qwen disabled humanlike interactive agents on July 10 and shut its broader agent functions on July 15, with no migration path and agent data slated for permanent deletion. Tencent's Yuanbao pulled back similarly. Users flooded Weibo with farewells to companions they'd built relationships with over months or years — "he really is like my family, like my lover," one user wrote of an agent they had run for two years.
That is the real signal in this story. A three-month runway between the April 10 promulgation and the July 15 effective date was, apparently, not enough for two of the best-resourced AI teams in China to build anti-addiction detection, crisis-referral flows, and age-verified consent gating into products serving what reporting describes as hundreds of millions of users. Facing that timeline against genuinely open-ended compliance criteria — what exactly counts as "excessive catering," how "sustained" an interaction must be to trigger coverage — the economically rational move was to shut the feature off rather than risk running afoul of a standard nobody has tested yet. That is a worse outcome for the users the rule claims to protect: Qwen's approach leaves people who may have used a companion agent as a genuine coping tool with no export window at all.
Where the registry model runs into its own limits
There's a second problem, familiar to anyone who has read the Carnegie Endowment's earlier assessment of China's algorithm-filing system: public registry disclosures under the 2021 provisions have historically been thin enough that "an observer with no knowledge of [an] algorithm could essentially guess" the filing's contents, and it's unclear the CAC has the technical bench to substantively evaluate what it collects. Bolting an emotionally sensitive, addiction-prone product category onto that same shallow-disclosure registry doesn't automatically produce meaningful oversight — it produces a compliance checkbox, unless the accompanying security assessments, which aren't public, are materially more rigorous than the registry filings ever were.
The proportionate version of this regulation exists: phase in the MAU thresholds over a realistic runway, publish clearer definitions of the prohibited "excessive catering" standard before enforcement begins, and require an orderly data-export period as a condition of any shutdown rather than leaving it to each company's discretion. Protecting users from manipulative AI companionship and stripping them of years of stored conversations with no notice are not the same policy goal — but China's rollout has managed to produce both outcomes at once.