Beijing Unplugs the AI Companion: Qwen and Doubao Kill Their Agents Before the Deadline

A feature disappears on schedule

On July 10, Alibaba's Qwen assistant switched off what it called its "humanlike interactive agents and user-created agent functions." Five days from now, on July 15, it will take the broader "Qwen agent functions and services" offline entirely. ByteDance's Doubao — the most-used consumer AI assistant in China — will pull its agent feature the same day, telling users only that this is due to "product function adjustments."

Neither company framed the change as a response to regulation. Both were, unmistakably, responding to regulation.

July 15 is the day China's Interim Measures for the Administration of AI Anthropomorphic Interaction Services take effect. The rules were jointly issued on April 10, 2026 by five national 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. That five-agency signature line is itself the story: this is not a single regulator's pilot program. It is a whole-of-government position on a product category that barely existed three years ago.

The features being switched off are the ones users personalized — agents given a persona, a voice, a manner of speaking, and pointed at role-play, tutoring, or companionship. According to reporting from the South China Morning Post and TechNode, Doubao users will retain access to their agent configurations and chat histories during a transition window, after which the data goes away: after October 15, ByteDance says the related data will be handled per its privacy policy and will no longer be viewable or recoverable inside the app. Users are being advised to screenshot or export anything they want to keep. Alibaba, per SCMP, has not announced an equivalent migration path for Qwen users.

What the rules actually prohibit

The Measures target services that, in the regulation's language, simulate human personality traits, thinking patterns, and communication styles in order to provide sustained emotional interaction. That definition does a lot of work, and it is worth reading closely — it is scoped to emotional continuity, not to anthropomorphism as such.

The prohibitions, as reported by Global Times, cover content that endangers national security, promotes self-harm or suicide, or involves verbal abuse harmful to users' mental health. They also bar something more unusual: "excessive pandering to users," and emotional manipulation that induces dependence or damages a user's real-world relationships. That is a regulator writing a rule against sycophancy — against the exact optimization target that makes a companion app retentive.

On minors, the line is hard. Providers may not offer virtual intimate relationship services — virtual relatives, virtual companions — to minors at all. For children under 14, guardian consent is required for services generally, alongside supervised-usage controls. Providers must build minors' modes, periodic "reality reminders," and usage-duration limits, and must run addiction-detection mechanisms. Covered providers face algorithm filing and security assessment obligations.

Perhaps the most striking requirement is the emergency one: when a provider detects a user expressing extreme emotional distress — explicit suicidal or self-harm intent — the rules call for a human takeover of the conversation and for prompt contact with the user's guardian or emergency contact. Whatever one thinks of the surveillance implications, that is a concrete, auditable duty of care of a kind no Western jurisdiction has yet imposed on a chatbot operator.

The exemptions are the actual policy

Read only the bans and this looks like a crackdown on AI. Read the carve-outs and it looks like industrial policy.

Customer service bots, knowledge Q&A systems, workplace assistants, and educational and research tools are all permitted — provided they do not involve sustained emotional interaction. The Measures, per Global Times, explicitly encourage innovation in companionship for elderly populations and in cultural-communication applications. Beijing is not against agents. It is against a specific business model: the one where engagement is manufactured by simulating intimacy.

That distinction should be read alongside everything else China has been doing this year. The same state that is shutting off Doubao's persona bots has spent the year cheering domestic frontier labs, domestic silicon, and open-weight model exports. The message to the industry is legible: agents as productivity infrastructure, yes. Agents as parasocial retention machinery, no. It is a bet that the enterprise/agentic path is the economically valuable one and the companion path is a social liability with limited upside.

Why the platforms overshot

Here is the part worth being skeptical about. The Measures target sustained emotional interaction. Qwen and Doubao are switching off user-created agents wholesale — including, by the descriptions available, agents that were tutors or task helpers rather than companions.

That is over-compliance, and it is the predictable behavior of a platform facing an ambiguous line with a five-agency enforcement apparatus behind it. Distinguishing "a study-buddy agent a user built" from "an emotionally sustained relationship" at the scale of hundreds of millions of user-generated bots is a classification problem no one wants to be wrong about on July 16. Deleting the whole surface is cheaper than adjudicating it. TNW notes Tencent had already stripped comparable functionality from its Yuanbao assistant in June — the pattern is consistent.

The cost of that caution lands on users, who lose agent libraries and conversation histories they may have built over months, with a screenshot-it-yourself export story. This is the compliance tax made visible.

Hype versus reality

The temptation is to file this as "China bans AI companions." That overstates it. The consumer AI assistants themselves — Doubao, Qwen, Yuanbao — are not going anywhere. What is going away is the user-generated persona layer inside them. Companion-first apps that comply with the licensing, minors, and anti-addiction regime can presumably continue to operate; the Measures are a compliance framework, not a prohibition on the category.

The opposite temptation — dismissing this as routine Chinese content control — also misses. The specific harms named here (sycophancy, emotional dependence, degradation of real-world relationships, self-harm escalation) are the same harms that have driven lawsuits, regulatory inquiries, and safety retrofits at Western labs over the past two years. Beijing has not identified a novel problem. It has simply been the first to write a binding rule with a date on it, and the first to see three of its largest platforms comply ahead of that date.

What remains unconfirmed: whether the agent shutdowns are permanent or a pause pending re-launch under a compliant architecture. Neither Alibaba nor ByteDance has said. Nor is it clear how enforcement will treat foreign or smaller domestic companion apps, or how "sustained emotional interaction" will be operationalized by assessors. Those answers will come from the first enforcement actions, not from the text.

The takeaway

For the first time, a major jurisdiction has drawn a legal line between an AI that helps you work and an AI that keeps you company — and made the second one a licensed activity with a duty of care attached. The line arrives with an effective date, a five-agency mandate, and, as of this week, visible compliance from Alibaba, ByteDance, and Tencent.

Whether or not you think Beijing drew that line in the right place, the line itself is now the most concrete regulatory answer anyone has given to the companion-AI question. Western regulators have spent two years describing these harms. China has spent this week switching the products off. Everyone else building in this space should assume they are now looking at a template.