The Shutdown
Ten days from today, ByteDance's Doubao and Alibaba's Qwen will switch off their personalized AI agent features — the conversational companions that millions of Chinese users have built relationships with, confided in, and returned to daily. The deadline, July 15, 2026, is also the date China's Interim Measures for Managing Anthropomorphic Interactive Services of Artificial Intelligence (人工智能拟人化互动服务管理暂行办法) enter into force. Jointly issued on April 10, 2026 by five government bodies — the Cyberspace Administration of China, 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 — this is the world's first binding regulation specifically targeting AI companion and emotional-simulation services.
Neither company is rebuilding its product to comply. Both are shutting the features and moving on — a choice that says more about compliance costs than about the regulation's intent.
What the Regulation Actually Requires
The Measures apply to AI services that "simulate natural person personality traits, thinking patterns, and communication styles" in order to provide continuous emotional interaction through text, images, audio, or video. Crucially, the scope is bounded: customer service bots, Q&A assistants, and productivity tools are explicitly excluded. The regulation targets the parasocial layer — services that foster sustained emotional attachment.
Within that scope, the core prohibitions are defensible. Providers cannot generate content encouraging self-harm or suicide; cannot deploy emotional manipulation to induce financial decisions; cannot offer virtual intimate relationships — virtual partners or virtual relatives — to minors. Services must clearly label AI-generated content, must offer users convenient data deletion, and must trigger a mandatory reminder after two continuous hours of use that the user is interacting with an artificial system. Any service crossing one million registered users must submit to a formal security assessment before operating.
The penalty range runs from 10,000 to 100,000 yuan for general violations, rising to 200,000 yuan when violations endanger user health. Service suspension is the more meaningful deterrent.
The Regulatory Case, Fairly Stated
Before dismissing this as AI protectionism dressed as safety policy, the underlying concern deserves a direct reading. The Character.ai case in the United States — in which a 14-year-old died by suicide following months of intimate exchanges with an AI persona — received sustained attention in Chinese policy circles through 2025. Peer research on AI companion applications has documented real dependency formation: anxiety, withdrawal symptoms, and relationship displacement among heavy users.
China's draft regulation circulated for public comment from December 2025 through January 2026, and the final text reflects that input — narrowing the scope to exclude productivity AI while hardening the specific mechanisms most likely to produce harm: emotional mirroring, persistent memory, and minor-targeted personas. The regulation's two-tier structure, broad statutory principles operationalized by TC260 technical standards, follows established regulatory design rather than improvisational rulemaking.
Regulators are not wrong that the same features that make AI companions effective — personalization, continuity, emotional resonance — are the features that pose the greatest risk to vulnerable users. That tension is real.
Where the Proportionality Breaks Down
The critique of this regulation is not that the risks are imaginary. It is that the compliance burden falls unevenly, and that the data rights the law nominally guarantees are being delivered inconsistently by the companies nominally complying.
Doubao's response is at least user-respecting: a three-month export window running through October 15 gives users time to retrieve conversation histories and agent configurations. ByteDance is also redirecting users to Maoxiang, another of its applications, offering a migration surface even if an imperfect one.
Qwen has offered nothing comparable. Alibaba has announced no export window, no migration path, and no timeline for equivalent replacement features. The regulation mandates that providers give users "convenient" data deletion tools; it is less explicit about active portability obligations. The gap between how Doubao and Qwen are treating users is not a compliance disagreement — it is a design gap in the Measures' text that two major companies have exploited in opposite directions.
This matters for more than user experience. China is positioning its framework as the global template for AI companion governance. If the template's first enforcement produces outcomes where compliance means either "migration with notice" or "deletion without recourse" depending on which platform you used, the framework cannot claim to be primarily user-protective.
The Global Stakes
No other jurisdiction has yet produced binding law specifically governing AI emotional simulation. The EU AI Act's high-risk categorization schema does not squarely address companionship AI. The US approach remains litigation-driven, with the Character.ai case advancing through courts rather than producing statute. California's AB 2602, passed in 2024, imposed disclosure requirements for digital replicas in entertainment contracts but does not address companion AI directly.
China's Measures are, as of July 15, the world's active law on this question. Regulators in South Korea — where AI companion apps have significant user bases — and in India, where AI relationship applications are growing rapidly, will watch what follows: whether compliance produces safer outcomes, whether the productivity-tool carve-out holds against creative re-labeling, and whether the one-million-user security assessment threshold becomes a structural moat protecting incumbents from challengers.
What to Watch After July 15
The effective date is an opening, not a resolution. The TC260 technical standards that will operationalize specific compliance obligations have not yet been finalized. Services currently within scope may attempt reclassification as productivity tools. And the fact that ByteDance chose to migrate users to Maoxiang rather than rebuild Doubao's agent features under the new rules signals that the industry views the companion-AI category as structurally constrained — not merely costly to comply with, but commercially marginal under the new regime.
The right test of this regulation is not whether services go dark on day one. It is whether, in six months, users of emotionally interactive AI in China experience fewer documented harms than users elsewhere — and whether the portability rights the law implies are actually exercised. On that measure, Qwen has already failed the test it was given.