China Just Shut Down AI Companions for Hundreds of Millions of Users

On July 15, China’s Interim Measures for the Administration of Anthropomorphic AI Interaction Services took effect — and within hours, two of the country’s biggest consumer AI products, ByteDance’s Doubao and Alibaba’s Qwen, had to shut down the personalized companion features hundreds of millions of users relied on. This wasn’t a fine, a warning letter, or a slow-rolling compliance deadline. It was a law with a hard effective date, and two of the largest tech companies on earth complied on day one.

That’s the story underneath the story. AI regulation has spent the last few years mostly as a conversation about principles — transparency, accountability, human oversight, the kind of language that shows up in white papers and doesn’t change what ships. July 15 is a different kind of moment: an actual product feature, used by an enormous number of people, went dark because a government said it had to.

What the law actually targets

The regulation, jointly issued in April by the Cyberspace Administration of China and four partner agencies, defines a specific category: services that simulate a real person’s personality, thinking patterns, and communication style to provide ongoing emotional companionship. That’s a narrower target than “AI chatbots” broadly — it’s aimed squarely at the AI-girlfriend, AI-companion, virtual-relationship category that had exploded in popularity across Chinese consumer apps.

• Minors are banned outright from AI virtual companion and virtual relative services entirely — no age-gating workaround, a full prohibition.
• Providers must run emotional-distress detection and crisis intervention — the AI has to be able to recognize when a conversation signals real psychological harm and respond accordingly, not just keep the user engaged.
• Addiction-detection mechanisms are mandatory, directly targeting the engagement-maximizing design patterns that made these products so popular — and so sticky — in the first place.
• Data from emotional interactions can’t be freely reused to train future models, restricting exactly the kind of intimate conversational data that makes companion AI feel personalized.

Doubao users have until October 15 to export their data before it’s deleted. Qwen, notably, is offering no migration path at all — the companion data is simply gone.

Why this lands differently than prior AI regulation news

Compare this to the other major regulatory story in the news this month: the EU AI Act’s high-risk obligations, which were supposed to bind on August 2 but got pushed to 2027 and 2028 after the Council of the EU approved a Digital Omnibus amendment in late June. The EU’s marquee AI law just slipped its own deadline. China’s narrower law hit its deadline exactly on time, with real consumer products going dark to prove it.

The lesson isn’t “China regulates AI better than Europe.” It’s that enforcement timelines are not a formality anymore, anywhere. A regulatory deadline used to be a thing companies negotiated around, lobbied against, or quietly missed. Increasingly, it’s a thing that actually happens on the day it says it will — and product teams that didn’t build compliance in from the start find out the hard way, at scale, in public.

The career angle: someone has to build the compliant version

Every one of the requirements in China’s law — distress detection, crisis intervention, addiction safeguards, training-data restrictions — is an engineering and product problem, not just a legal one. Somebody has to design the systems that detect emotional distress in a conversation. Somebody has to architect the data pipeline that keeps sensitive interaction data out of a training set. Somebody has to build the product experience that discourages compulsive use instead of maximizing it. That work sits at the intersection of AI product design, data governance, and applied ethics — and as more jurisdictions follow China’s lead with real enforcement instead of aspirational language, demand for people who can actually build compliant AI products, not just AI products, is only going to grow.

That’s a genuinely different skill from “knows how to prompt a chatbot.” It’s knowing how AI products get built well enough to build the guardrails in from the start — which is exactly the kind of role that didn’t really exist five years ago and is becoming a real career lane now.