China Just Made Emotional AI Illegal. Doubao and Qwen Ripped Out Their Companion Agents Overnight.

TL;DR

As of today, China regulates the relationship your AI has with a user, not just what the model can do. The Cyberspace Administration of China and four other agencies put the Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services into force on July 15, 2026, banning AI built to simulate a human personality and sustain an emotional bond. To comply, ByteDance's Doubao and Alibaba's Qwen yanked their custom-agent features. Millions of user-made companions, tutors, and role-play personas went dark, and the exact "give your agent a persona" pattern that Western labs ship by default is now illegal there.

What the law actually bans

The measures were issued in April 2026 and took effect today. They target, in the regulation's own words, services that "simulate human personality traits, thinking patterns and communication styles to provide sustained emotional interaction." That last phrase is the whole ballgame. The trigger is not the size of the model or the cleverness of the answer. It is whether the product is designed to become someone you feel attached to.

Think of it less like banning a specific tool and more like a rule for a bartender: still allowed to pour your drink, no longer allowed to become the person you tell your problems to. Capability is fine. Companionship is the regulated act.

That is why the carve-outs matter. The rules explicitly exempt customer service bots, knowledge Q&A, workplace assistants, and education or research tools, as long as they do not cross into sustained emotional interaction. You can still ship a bot that answers questions. You just cannot ship one that misses you.

What builders lost overnight

The two products that moved first are not fringe apps. Doubao and Qwen are among China's most-used AI services, and both offered the same feature: spin up a named assistant, tutor, role-play character, or companion with a fixed persona and tone. That is precisely the category the law defines, so precisely the category that had to go.

Alibaba moved early, disabling Qwen's humanlike interactive agents and user-created agent functions on July 10, then shutting broader agent services on the 15th. ByteDance disabled Doubao's custom-agent creation on the 15th. New agents cannot be made; existing ones stopped responding.

The data endings differ, and if you build on either platform this is the part to read twice. Doubao users keep read-only access to their agent configs and chat histories until October 15, 2026, after which the data is handled under ByteDance's standard privacy policy and is no longer recoverable in the app. Qwen users got no comparable grace period: Alibaba has said agent configurations and conversation histories will be permanently deleted, with no announced migration path.

The rules under the ban

The measures are not only a ban list. For services that stay inside the lines, they impose a compliance stack: interactions must be clearly disclosed as AI, providers must run anti-addiction friction and usage-time notifications, and there are security assessments plus provincial filing obligations before you ship. The stated worries read like a companion-app risk register: extremist ideas, privacy leaks, harm to physical and mental health, and dependence or addiction.

If you have built for consumers in China, none of that is a surprise in kind, only in specificity. Beijing has now written "your chatbot must not be too good at being a friend" into administrative law, with filing paperwork attached.

Why this matters beyond China

Here is the reversal a Western builder should sit with. The feature China just outlawed is a default in the rest of the market. Persona presets, named assistants, memory that makes the bot feel like it knows you, companion apps like Replika and the character bots on Character.AI: that is not an edge case abroad, it is the product. One government just declared the whole pattern a regulated activity, and two of the largest AI apps on earth complied in under a week.

You do not have to think Beijing got the balance right to notice the signal. Emotional stickiness is the growth lever half the consumer-AI industry is pulling, and it is now the exact surface a regulator chose to grab first. Anti-addiction limits, mandatory "you are talking to an AI" disclosure, and hard data-deletion deadlines are a concrete template, and templates travel. If you are shipping anything with a persona and a memory, the compliance question is no longer hypothetical; it just has a jurisdiction and a date attached.

The near-term takeaway for anyone with users in China is duller but sharper: audit whether your product's charm is legally a liability there, and assume your happiest retention metric is the one an auditor circles first.

Key Takeaways

• The law: China's Interim Measures for Anthropomorphic AI Interaction Services took effect July 15, 2026, banning AI designed to simulate a personality and sustain an emotional bond.
• The trigger: "Sustained emotional interaction" is the legal line. Customer service, Q&A, workplace, and education tools are exempt if they stay on the capability side of it.
• The casualties: Doubao and Qwen disabled custom-agent creation; Qwen cut humanlike agents on July 10, both stopped agent creation on the 15th.
• The data: Doubao gives read-only access until October 15, then the data is unrecoverable. Qwen announced permanent deletion with no grace period.
• The compliance stack: AI disclosure, anti-addiction friction, usage notifications, security assessments, and provincial filing for services that remain.
• The signal: The exact companion pattern Western labs ship by default is now illegal in one major market. Emotional stickiness is the first surface a regulator reached for.

Sources: TechNode, South China Morning Post, Crypto Briefing, AI Weekly