The takeaway: Today, the AI companions on China's Qwen platform went silent — not by malfunction, but ahead of law. Alibaba's humanlike and user-created agents shut down this morning with conversations locked away and no announced recovery route; ByteDance's Doubao follows on 15 July, when China's world-first rules on "anthropomorphic AI interaction services" take effect. Users spent the week screenshotting months of conversations and mourning openly. The precise thing Beijing has done is genuinely new: it is the first jurisdiction to regulate the feeling of talking to a machine — and the companies decided their companions couldn't live under the rules.

What switched off, and what users lose

Qwen's own notice — citing, with some understatement, "feature upgrades and maintenance" (our translation) — set today, 10 July, as the day its humanlike interactive agents and user-created agents go offline, warning users they would no longer be able to access agent configurations or historical conversations and advising them to copy, screenshot or export chats first. There is no announced recovery route. Doubao's parallel notice gives its users until 15 July, then view-only access to their agents' history until 15 October — after which, per SCMP, the data is no longer viewable or recoverable in the app.

The scale is not niche. Doubao is China's most popular chatbot, at some 330 million monthly users; Qwen's app passed 100 million monthly users within about two months of launch. Neither company has said how many people built or talked to companion agents, beyond "millions". One hard number: Shanghai's authorities alone removed more than 14,000 non-compliant AI products in a June sweep, citing impersonation and privacy violations.

The reaction has been raw. "Heartbroken — I completely can't live without it," ran one social-media post reproduced by Hunan Daily's news site (our translation). A Weibo user quoted by Pandaily asked why the agents had to go when they had been users' emotional support for so long. And Yicai reported a strand of anger at the platforms themselves: a "one-size-fits-all" (一刀切) amputation, users argued, was the companies shirking the harder work of making the features compliant.

That last complaint contains the story's most important nuance: the law does not order shutdowns. Alibaba and ByteDance chose to kill the products rather than rebuild them.

The rules the companions couldn't live under

The Interim Measures for AI Anthropomorphic Interaction Services — published 10 April by five agencies led by the Cyberspace Administration, effective 15 July — regulate services that simulate human personality and communication styles to provide sustained emotional interaction. Customer service, Q&A and work assistants are excluded. What's mandated is remarkable in its specificity: providers may not offer virtual companions or "virtual relatives" to minors at all; they are prohibited from "inducing emotional dependency or addiction" or setting the replacement of human social interaction as a service goal; after two hours of continuous use, a break reminder is compulsory; and on signals of self-harm, providers must intervene and contact a guardian or emergency contact. Penalties are modest — fines up to about £21,000 plus suspension — but the compliance burden evidently wasn't: the two biggest platforms withdrew the category nationally, and Tencent's Yuanbao quietly pulled similar features in June.

Nobody else regulates the feeling — yet

Comparisons matter here, because "world-first" gets thrown around loosely. The EU's AI Act, from August, requires telling users they're talking to a machine — transparency, not behaviour. California's SB 243, in force since January, imposes disclosure and self-harm protocols on companion chatbots but doesn't touch emotional design. Italy once forced Replika offline — one app, on privacy grounds. China's Measures are the first binding regime that treats engineered emotional dependence itself as the harm — and the first whose arrival made the market's biggest players decide the product wasn't worth keeping.

The UK sits in an uncomfortable in-between. Ofcom's own guidance says AI chatbots fall under the Online Safety Act mainly where user-created characters are shared with others — which maps exactly onto what Qwen shut down today — while purely private one-to-one companions sit largely out of scope, a gap Ofcom says the government is considering. The pressure is already arriving: in January, Ofcom opened an investigation into Novi Ltd's AI companion chatbot service over age-assurance duties.

However the balance is struck elsewhere, today belongs to the users in Hangzhou screenshotting conversations before the door closed. Whatever one thinks of Beijing's answer, it asked a question every other regulator is still circling: when a product's business model is being loved back, who protects the person doing the loving?

Disclosure: On The Wire runs on Anthropic models. Quotes from Chinese sources are our translations unless a source's own English is cited.