A world-first regime, and an immediate casualty count
On July 15, 2026, China's Interim Measures for the Administration of AI Anthropomorphic Interactive Services (人工智能拟人化互动服务管理暂行办法) took effect, becoming the first binding national framework anywhere aimed specifically at AI companions and emotionally interactive chatbots. The rule was jointly issued on April 10, 2026 by five agencies — the Cyberspace Administration of China (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 — following a draft that CAC circulated for public comment in December 2025 (CAC announcement).
The effect was immediate and visible. ByteDance's Doubao and Alibaba's Qwen — China's two largest consumer AI assistants — simultaneously shut down their persistent AI-agent and companion-creation features rather than attempt to redesign them for compliance. Doubao is giving users read-only access to old chat data until October 15, 2026, after which it becomes unrecoverable; Qwen announced no migration path at all (TechNode). That two of the best-resourced AI companies in the world's second-largest tech market chose deletion over adaptation is itself a data point: it suggests the rule's design constraints, not just its compliance costs, are hard to reconcile with how companion products are built.
What the rule actually requires
The Measures apply narrowly — to "sustained emotional interaction services" that simulate a natural person's personality and communication style — explicitly carving out customer service bots, Q&A tools, and productivity assistants. Within that scope, providers must: continuously and prominently disclose that the user is interacting with AI rather than a human; file their algorithms and undergo security assessment; deploy anti-addiction reminders after roughly two hours of continuous use; detect signs of emotional over-reliance or self-harm risk and escalate to intervention protocols; and, most strikingly, bar minors entirely from "virtual intimate relationship" features — virtual romantic partners or virtual family members (CAC Q&A). The Measures apply to any service "provided to the public within the territory of the People's Republic of China," a formulation legal analysts read as covering offshore providers serving Chinese users even without the explicit extraterritorial trigger found in China's 2023 Generative AI Measures (IAPP).
Steelmanning Beijing's case
The strongest argument for the rule is not hypothetical. Regulators worldwide — including US state legislatures and the FTC — have opened inquiries into AI companion apps following reported cases of minors forming unhealthy attachments to chatbots, including instances linked to self-harm. A mandatory non-human disclosure, a hard age gate on intimacy-simulating features, and a duty to detect and interrupt addictive usage patterns are not exotic asks; they resemble consumer-protection guardrails already applied to loot boxes, autoplay, and other engagement-optimized product design. China got here first, but the underlying concern — that emotionally responsive AI is optimized for retention in ways that can be actively harmful to vulnerable users, particularly minors — is one that serious, non-authoritarian regulators share.
Where the rule overreaches
The problem is not the goal but the mechanism. By mandating specific interaction-level interventions — a two-hour reminder threshold, dynamic AI-disclosure pop-ups, guardian-monitoring dashboards — rather than outcome-based standards, Beijing has effectively told an entire product category it must re-architect around state-specified friction points or exit the market. Doubao and Qwen didn't choose deletion because compliance was impossible; they chose it because the compliant version of a persistent-memory companion is a fundamentally different, less sticky product, and neither company judged that product worth building on a compressed timeline. That is a legislature substituting its product-design judgment for the market's, and the two companies affected serve a combined user base in the hundreds of millions — Doubao alone reported 345 million monthly active users in its most recent disclosure (TechNode).
The extraterritorial ambiguity compounds this. Because the Measures apply to services reaching Chinese users without clearly excluding offshore providers, any global AI company with Chinese-market ambitions now faces a Beijing-shaped design constraint before it ever launches a companion feature elsewhere — a soft form of regulatory export that other jurisdictions watching this rollout should not import wholesale.
The proportionate alternative
A narrower rule — mandatory human-in-the-loop crisis escalation, verified age gates for intimacy features, and clear AI-disclosure requirements, enforced against outcomes rather than dictating interface mechanics — would have addressed the genuine minor-safety concern without forcing wholesale product shutdown. The EU's approach under the AI Act, which requires disclosure of AI interaction without prescribing session-length interventions, is closer to that model. As US states and the EU begin drafting their own AI-companion rules this year, Beijing's experiment is worth studying for its stated goals — and avoiding for its method.