Regulation / Jul 6, 2026 / 4 min
Beijing Banned AI Lovers, Kept the Office Bots
On July 6, ByteDance's Doubao and Alibaba's Qwen began shutting down customizable AI companion personas ahead of China's first national anthropomorphic-AI law taking effect July 15 — deleting chat histories for hundreds of millions of users while leaving workplace agents untouched, nine months before Washington finished arguing about voluntary frontier-model previews.
ByteDance and Alibaba are deleting the companion layer from China's two biggest consumer AI apps — not because the models failed, but because Beijing's first national law on anthropomorphic AI takes effect July 15 and neither platform could comply in time.
What's new: Doubao and Qwen notified users over the July 4–5 weekend that customizable agent personas — named tutors, role-play characters, and emotional companions — are going dark. Workplace bots, Q&A, and customer-service agents stay.
The shutdown calendar:
- Qwen: Humanlike and user-created agents disabled July 10; all agent functions offline July 15. Prior conversations become inaccessible the same day.
- Doubao: Agent feature offline July 15. Chat histories remain read-only until October 15, then are wiped and unrecoverable inside the app.
- Tencent Yuanbao pulled a similar persona feature in June. NetEase is aligning too, per Nikkei.
The law behind it: Five Chinese ministries — led by the Cyberspace Administration — issued the Interim Measures for the Administration of AI Anthropomorphic Interaction Services on April 10. It takes effect July 15.
The rules govern services that "simulate human personality traits, thinking patterns and communication styles to provide sustained emotional interaction," per the South China Morning Post.
They explicitly exclude customer service, knowledge Q&A, workplace assistants, and education tools that don't involve sustained emotional bonding.
What Beijing fears:
- Extremist content and privacy leaks
- Mental-health harm and unhealthy emotional dependency
- Minors forming virtual intimate relationships
Providers crossing 100,000 monthly active users must file safety assessments. After two hours of continuous use, platforms must prompt breaks. Self-harm signals trigger mandatory human intervention.
The scale: Doubao had 345 million monthly active users as of March, per QuestMobile data cited by The Straits Times. Qwen had 166 million. ByteDance is spending roughly 200 billion yuan on AI this year.
Both companies spent 2026 racing to own China's default assistant. Now they're ripping out the feature users actually bonded with.
Users are mourning: One Weibo poster called the agents "long-standing emotional support" and complained there's no way to export chat histories, per The Next Web. Doubao is directing users to screenshots.
Pan Helin, an expert on China's Ministry of Industry and Information Technology committee, told the SCMP the policy prioritizes "safety, practical use, and standardisation" — and that "current agents are not yet mature."
The split with the West: This is the world's first dedicated national rule for AI companionship. Washington still has no federal companion-AI law.
- California SB 243 (effective January 2026) mandates disclosure and crisis hotline referrals.
- Character.AI settled multiple teen-suicide lawsuits in January 2026, then banned minors entirely — but kept the engagement architecture for adults.
- Forbes estimates the global companion market at roughly $30 billion, with Character.AI reporting 233 million registered users.
Beijing isn't waiting for litigation. It's regulating the relationship structure — not just downstream harm.
Why it matters for the agent race: China issued agent-development guidance in May and national agent-interoperability standards in June. The pattern is clear: agents as infrastructure, companions as liability.
Silicon Valley is pitching the same stack — personalized agents inside every app — while Meta tests rival chatbots on sensitive teen topics and OpenAI gates GPT-5.6 Sol behind government previews. Beijing just proved you can have an agent economy without letting users fall in love with the product.
Convina's view: This is the first time a major AI power regulated attachment itself — not content, not exports, not kill switches on frontier models. The West will call it censorship. Investors should read it as product segmentation: China is building worker agents at national scale and deleting the companion layer that Character.AI built its entire valuation on. If your agent strategy assumes emotional lock-in, Beijing just showed the exit ramp — and 345 million users are being pushed through it next week.