Doubao and Qwen turned their companion agents off on July 15, 2026 rather than comply with China's new companion-AI law. The law regulates the emotional bond a companion creates — not the model behind it — and you cannot bolt "please don't get attached to me" onto a product whose entire value is attachment. So two of the most-used apps in the world's largest AI market judged a fully compliant companion worse than no companion at all, and pulled the feature outright.

Why a founder outside China should care: this is the first regulation anywhere to make "is it a tool or a companion?" a legal question — and it's the exact line most likely to reach your market next. If your product holds a persistent persona or an emotional hook, the classification below is coming for you.

Here's the whole story in one screen — the tool-vs.-companion split, why it forced an off switch, and why it travels:

The durable takeaway isn't the shutdown — it's the classification test underneath it, which is where this piece ends up. First, what actually happened.

Update, Aug 19, 2026: the Western platforms are arriving at the same line from the other direction. On Aug 18, OpenAI made a locked-down "ChatGPT for Teens" the default for any account it predicts is under 18, banning romantic language and instructing the model not to imply it has feelings — self-regulating the same surface under legal pressure that China regulated by law.

The law regulates a relationship#

The Interim Measures for the Administration of Anthropomorphic AI Interaction Services, released by Chinese regulators on April 10 and effective July 15, do not read like other AI rules. They are not about model weights, training data provenance, or output safety in the usual sense. They are about what happens when software pretends to be a person.

The provisions are specific. A service must display a conspicuous alert that the user is talking to an AI and not a natural person. When it detects signs of over-dependence — or simply on a new login — it must throw a pop-up reminding you that what you are reading was generated by a machine. After two continuous hours of use, it must prompt you to stop. It may not offer virtual-companion or virtual-relative services to minors at all, and other anthropomorphic services to under-14s require a guardian's consent. And if a user signals self-harm or extreme distress, the provider must intervene — surface help, contact a guardian or emergency contact.

Read those as a set and a shape emerges. Every requirement is aimed at the bond, not the capability. None of it constrains what the model can compute. All of it constrains what the model is allowed to be to you.

Western AI law regulates what a model may do. China just regulated what a model may be.

Why both giants chose the off switch#

Here is the part worth sitting with. ByteDance and Alibaba are not small shops that lacked the engineering budget to comply. They are among the most capable AI organizations on Earth, and they had three months' notice. They chose to kill the feature rather than adapt it.

That is the signal. Compliance here is not a filter you add to a pipeline. The mandated behaviors are in direct tension with the product's reason to exist. A companion agent is valuable precisely because it sustains a believable, consistent persona and an emotional relationship across sessions. A companion that must periodically break character to insist it is not real, that cannot be a companion to a minor, and that is legally obligated to puncture the very dependency it was optimized to create is — functionally — not a companion. You cannot retrofit "please don't get attached to me" onto a product whose entire value is attachment.

So the math both companies apparently ran was blunt: a fully compliant companion is a worse companion than no companion, and the liability of getting the crisis-intervention and minors provisions wrong is severe. Switching it off was the cheaper failure.

The category split nobody has named yet#

For everyone building agents outside China, the temptation is to file this under "Chinese regulation, doesn't apply to me" and move on. That misses the durable idea.

Until now, "AI agent" has been one regulatory blob, sorted mostly by capability and risk — what the EU AI Act tiers, what US executive orders enumerate. China, which has otherwise been busy standardizing how agents interconnect, has just drawn a different line, and it runs across that blob rather than along it. On one side: tool agents — things that book, retrieve, summarize, execute, and never pretend to be anyone. On the other: persona agents — things whose value is a simulated self and a relationship. The measures barely touch the first category and effectively regulate the second out of its most lucrative form.

That line is going to be copied, because it maps to a real harm the capability-based frameworks keep missing. The damage from a companion app is not usually a wrong answer or a biased output; it is a teenager who stops talking to humans. Regulators elsewhere are already watching the same wrongful-death lawsuits that pushed Beijing here. When they act, the first question they will ask a product is not "how capable is your model" but "is this a tool or a companion" — and every founder who has quietly shipped a "friend" while calling it an "assistant" will discover the distinction was load-bearing all along.

The models will keep getting better on both sides of that line. What changed on July 15 is that the line now exists.

Follow-up: the durable takeaway here isn't the shutdown — it's the classification test underneath it. We turned China's five requirements into a self-audit you can run on your own product today: tool or companion? here's how to tell which side of the line you're on. And for what the measures changed on the ground once they went live, see what founders actually do now that the rules are in effect.