On July 15, the shutdown that founders had been reading about since early July actually happened. ByteDance's Doubao and Alibaba's Qwen switched off their consumer AI-companion agents that morning, on the day China's new Interim Measures for the Administration of AI Anthropomorphic Interactive Services took effect. No last-minute reprieve, no quiet retrofit. The features went dark.
What is worth your attention a day later is not that they complied — we knew they would — but how differently they complied, and what the day revealed about where the regulatory line actually falls.
Two giants, two exits#
ByteDance and Alibaba did not do the same thing.
ByteDance redirected Doubao's users — roughly 345 million monthly actives — to Maoxiang, a separate ByteDance app, telling them they could create new agents and resume conversational services there. Doubao users keep read-only access to their old agent configs and chat histories until October 15, and the in-app notice tells them to save anything they care about the hard way: screenshots, or the text-share function. After October 15 the data falls under Doubao's standard privacy policy and is no longer recoverable in the app.
Alibaba's Qwen offered no such door. There is no announced migration path; agent configurations and conversation histories are simply being deleted.
ByteDance moved its companions to a building designed to pass inspection. Alibaba condemned the building.
The strategic read is clean. Doubao is a general assistant with hundreds of millions of users, and the mandated behavior — conspicuous "you're talking to an AI" disclosure, dependence-detection pop-ups, a two-hour break prompt, a hard ban on companions for minors, crisis intervention — cannot be sprinkled onto a general assistant without degrading it for everyone. So ByteDance did the sane thing: it moved the regulated product into a purpose-built container where the compliance architecture is designed in from line one, and kept Doubao itself clean. Alibaba, at least for now, chose to exit the consumer-companion category rather than build that container.
The detail founders should circle#
Here is the part that travels beyond China. The rules left enterprise, productivity, and customer-service agents untouched. The same week two of the world's largest AI companies killed their consumer personas, their productivity and enterprise agents kept running without a pause.
That is the whole point, and July 15 made it visible: the regulated surface is simulated personhood, not capability. Western AI law asks what a model may do — its risk tier, its outputs, its training data. China just enforced, in production, a rule about what a model may be to you. And the enforcement drew a bright, observable line between a tool and a companion.
Until this week, that line was a thought experiment you could wave away — the category split nobody had named yet, as we put it before the deadline. Now it is a compliance boundary with a body count of shipped features on one side and a business-as-usual on the other.
What to do with it#
If you build character, companion, or persistent-persona products — the kind where the value is the relationship — the Doubao/Maoxiang move is your template and your warning. Retrofitting disclosure and anti-dependence controls onto a product optimized for attachment tends to gut it; if that surface is your business, assume you will eventually need a compliance-native container, and design the disclosure and break-prompt behavior as first-class features, not afterthoughts.
If you build tools — agents that book, retrieve, transact, support, automate — the signal is quieter and more reassuring: the first real-world enforcement of "personhood" rules stepped around you entirely. Keep it that way. The cheapest way to stay a tool in a regulator's eyes is to not pretend to be a person in the first place.
Either way, "is this a tool or a companion?" is no longer a design question you get to answer with a shrug. Somewhere, it is already a line on a compliance form.
