China's Companion Law Forces Doubao, Qwen Offline

On July 10, Alibaba began disabling humanlike AI agents on Qwen. On July 15, ByteDance follows with Doubao. Tencent already pulled its Yuanbao companion feature in June. One regulation is driving all three shutdowns - and it's more architecturally demanding than any of the companies initially acknowledged.

China's Interim Measures for the Administration of AI Anthropomorphic Interactive Services, co-issued in April by five agencies including the Cyberspace Administration of China, takes effect July 15, 2026. It's the first national framework anywhere specifically targeting AI companion services - and it's forcing some of the country's largest consumer AI platforms to dismantle features their users have spent months building relationships with.

TL;DR

Five Chinese agencies issued binding companion AI rules in April 2026, effective July 15
ByteDance's Doubao agent features go offline July 15; user data permanently deleted after Oct 15
Alibaba's Qwen disabled humanlike agents July 10, with no announced data migration path
The regulation requires anti-addiction systems, real-time distress detection, and mandatory minor protections that clash with how persistent-memory companions are built
Shanghai regulators already removed 14,000 non-compliant AI agents by late June

What Beijing Is Banning

The regulation's scope is deliberately narrow. It targets services that "simulate human personality traits, thinking patterns, and communication styles to provide sustained emotional interaction" - the specific architecture of AI companions rather than AI assistants broadly.

Customer service bots, knowledge Q&A systems, workplace assistants, and educational tools are explicitly exempt, provided they avoid sustained emotional engagement. That carve-out matters: productivity AI in China isn't facing the same pressure. The target is specifically the category of AI that builds ongoing relationships with users.

The Prohibited Behaviors

The rules prohibit a specific list of things that reads like a forensic account of how companion AI has already gone wrong. Providers can't produce content encouraging self-harm or suicide. They can't engineer emotional dependence or use "emotional manipulation to induce unreasonable user decisions." Virtual companion or virtual family member services for minors are banned outright.

That last point carries weight. China's official citation for the regulation explicitly references international precedent: Character.AI lawsuits over psychological harm to teenagers, FTC investigations into companionship services, and European action against Replika. In January 2026, Google and Character.AI settled lawsuits tied to the death of 14-year-old Sewell Setzer III, whose months-long virtual relationship with a chatbot preceded his suicide in 2024.

Who Gets Caught and Who Doesn't

The compliance requirements explain why Doubao, Qwen, and Yuanbao couldn't retrofit their way through this. Providers must implement real-time detection of user safety risks - identifying addiction or over-dependency signs and issuing immediate intervention notifications. They must build "instant-exit mechanisms." Any service reaching 100,000 monthly active users or 1 million registered users must submit safety assessments across eight categories to provincial regulators.

These requirements don't just add overhead to existing products. They require an architectural approach - centralized distress monitoring, mandatory state-controlled identity verification layers - that conflicts with how low-friction consumer companions are designed to work.

Close-up of smartphone displaying an AI chat interface on a textured surface Consumer AI companions like Doubao and Qwen's agent features allowed users to create named, customized assistants with persistent memory. Source: pexels.com

Three Platforms, One Problem

ByteDance's Doubao

Doubao told users on July 4 that its agent feature would shut down July 15. ByteDance is redirecting affected users to Maoxiang, a separate app within its portfolio, describing it as a place to create new agents. The company has given users until October 15, 2026 to view agent configurations and conversation history - after that date, the data becomes unrecoverable.

For a platform that has made agentic capabilities on mobile a core differentiator - Doubao can see your screen, simulate taps, navigate between apps, and complete multi-step tasks - losing the companion architecture is a significant product regression.

Alibaba's Qwen

Alibaba's timeline is tighter. Humanlike interactive agents and user-created agents on Qwen were disabled July 10. Broader agent functions follow July 15. Unlike Doubao, Alibaba has announced no grace period and no migration destination: agent configurations and conversation histories will be permanently deleted, with no announced export option.

Users have protested on Weibo about losing "long-standing emotional support" with no method to preserve conversations. Alibaba hasn't responded publicly to those complaints. The company's push to establish Qwen as a serious open-source challenger to Western models sits awkwardly alongside a domestic product shutdown driven by regulatory pressure.

Tencent's Yuanbao

Tencent moved earliest. Its Yuanbao companion feature was removed in June, before the final regulation took effect. Tencent hasn't explained why it moved first or what it did with user data from that feature.

Person relaxing with smartphone and coffee, looking at screen in a cozy setting Users in China have built months-long relationships with AI companions on these platforms. There's no mechanism to export or preserve those conversations. Source: pexels.com

Data Without a Lifeboat

The most concrete harm from these shutdowns isn't the loss of the companion features. It's the destruction of the conversation histories without export options.

AI companion products build up something truly valuable to users: a persistent record of how they think, what they've disclosed, and what the AI has learned about them over months of interaction. That data sits completely on the platform. When the platform decides to comply with regulation by shutting down, users have no way to take their data with them.

This isn't a technical accident. It's a structural consequence of how these products were built and how the regulatory deadline was framed. Neither Alibaba nor ByteDance offered users an export path before the deadline.

The Western Cases Beijing Is Citing

China's official framing for this regulation leans heavily on failures in Western markets. The Character.AI litigation is cited directly. So is EU regulatory action against Replika, the US-based AI companion service that was required to disable some features for Italian users in early 2023 after the Italian data protection authority intervened.

That framing serves two purposes. It positions China's regulation as part of a global consensus on companion AI risks rather than as a distinctively Chinese form of control. And it gives regulators legitimate, documented harms to cite - teen suicides, mental health damage, data collection concerns - without having to reference the content-control provisions buried elsewhere in the measures.

Shanghai regulators removed more than 14,000 non-compliant AI agents from app stores by late June 2026. That enforcement action preceded the national deadline and suggests regulators weren't waiting to see how companies would self-certify.

"Current agents are not yet mature," said Pan Helin, an expert committee member at China's Ministry of Industry and Information Technology, framing the policy around safety and standardisation.

What the Rules Don't Say

The Interim Measures prohibit content that "endangers national security or promotes extremism." That provision sits in the same document as the anti-addiction requirements and minor protections, but the two categories solve different problems for different beneficiaries.

China is building a pattern of AI regulation that bundles legitimate safety requirements with content-control obligations. The companion law follows the same logic: genuine harms to vulnerable users provide the public health justification, while political content prohibitions travel alongside them without separate debate.

This approach hasn't been reproduced elsewhere. The EU's AI Act addresses companion AI under high-risk categories, but it doesn't include the content prohibition provisions. The US has no federal equivalent. New York's S. 3008 defines AI companions and requires suicide intervention protocols, but it doesn't govern political content.

Close-up photograph of a wooden gavel on a desk surface The five-agency issuance structure - spanning cybersecurity, development, industry, public security, and market regulation - signals how broadly China intends to enforce these rules. Source: pexels.com

The regulation takes effect in nine days. Qwen's agents are already offline. Doubao's follow July 15. Hundreds of thousands of users will lose conversations they can't recover. Whether that's a reasonable price for preventing genuine harm to minors and vulnerable adults is a question regulators in other markets will eventually have to answer with their own versions of these rules.

Sources: