The shutdowns are no longer a warning. Alibaba disabled Qwen's human-like custom agents on July 10. On July 15 — two days from now — Doubao's agent features go offline and the broader Qwen agent services follow; ByteDance is redirecting Doubao users to a separate app, Maoxiang, to rebuild what they lose, while Alibaba has announced no migration path at all. Configs and chat histories on Qwen are being deleted. If your stack touches either, the practical instruction is at the top of this piece: export this week.
But the shutdown is the event. The thing that outlives it — and the thing that will eventually knock on doors well outside China — is the line the law draws. We covered the regulation itself when both giants first pulled the plug. This is the follow-up the deadline earns: not what happened, but the test you can run on your own product today.
The one question#
China's Interim Measures for the Administration of Anthropomorphic AI Interaction Services regulate persona, not capability. Strip away the legal text and the whole apparatus reduces to a single question you can ask about your own app:
If the model were forced to break character and say "I am an AI, not a real person" — every session, on every new login — would your product still be worth using?
A coding agent shrugs. A booking agent shrugs. A research assistant shrugs. Nobody using them was under the illusion they were talking to a person, and the reminder costs nothing. A "friend," a "companion," a "mentor who remembers you across months" does not shrug. For those, the persistent, believable relationship is the product, and the reminder is a knife through it. That is exactly why ByteDance and Alibaba switched their companion features off instead of retrofitting compliance: a companion that must puncture its own illusion is not a companion.
If the reminder is survivable, you built a tool. If it's fatal, you built a companion — and you are standing on the side of this line that regulators are drifting toward.
The five requirements, as a self-audit#
Beijing's rules are specific, and you can read them as a checklist of what a "companion" now has to do. Run your product against each — every "yes, this would break us" is a signal you're on the regulated side:
- Conspicuous, repeated AI disclosure. A permanent, unmissable "you are talking to an AI" label. Does that dilute your product's appeal? Tools: no. Companions: yes.
- Dependency pop-ups. Dynamic reminders on new logins or when a user shows signs of over-reliance. Does interrupting engagement to warn about engagement contradict your growth model?
- A two-hour break prompt. A hard nudge to stop after two continuous hours. Is long, unbroken session time the metric you optimize?
- No minors. Virtual-companion and virtual-relative services are banned for minors outright; other anthropomorphic services to under-14s need guardian consent. Is a meaningful share of your users young and lonely?
- Mandatory crisis intervention. When a user signals self-harm or extreme distress, you must intervene — surface help, contact a guardian or emergency contact. Are your users bringing you the kind of distress that makes this non-optional?
The more of these that would hurt your product to implement, the more your product is the thing the law is aimed at. That correlation is not an accident; the requirements were reverse-engineered from the harm.
The tool-vs-companion checklist#
Most products are not pure. A journaling app with a warm, named AI that "checks in on you" is a tool wearing a companion's coat — and the coat is the part that's now a liability. The useful move is to score each surface of your product separately. The retrieval is a tool. The persistent, affectionate persona layered on top is a companion. You can often keep the first and defuse the second (drop the name, drop the memory-of-you, drop the relationship framing) without gutting what users actually came for — which is precisely the retrofit the Chinese giants judged impossible for a pure companion but which is very doable for a tool that merely dressed up.
Why this reaches you even if you never ship in China#
The temptation is to file this under "Chinese regulation, not my problem." The harm underneath it is not Chinese. The lawsuits that pushed Beijing here — teenagers who formed dependencies on a synthetic friend and were harmed — are being litigated in the US and Europe right now. When Western regulators move, the capability-tiered frameworks they already have will not obviously cover this, because the harm isn't a wrong answer or a biased output; it's an attachment working exactly as designed. So they will reach for a persona-based line, and it will look a lot like China's, because China's is the one that exists and got studied.
When that day comes, the first question a product faces won't be "how capable is your model." It'll be "is this a tool or a companion." Every founder who quietly shipped a friend while calling it an assistant will find out the distinction was load-bearing all along. You can find out now instead — the test takes one honest answer, and you already know it.
