On July 15, ByteDance's Doubao and Alibaba's Qwen switch off their companion and custom-agent features to comply with a new Chinese regulation. That shutdown has been covered as a China story — a quirk of Beijing's approach to AI. It isn't a China story. It's the third time in nine months that a major government has written "AI companion" into law as its own regulated category, and the first two were American.
Three laws, written apart, pointing at the same thing#
Rewind. New York's AI Companion Models law took effect on November 5, 2025. California's SB 243 — the first US statute to name "companion chatbots" — was signed on October 13, 2025 and came into force on January 1, 2026. China's Interim Measures for the Administration of Anthropomorphic AI Interaction Services land on July 15, 2026. Different legal systems, different drafters, no coordination. And yet, read side by side, they rhyme.
Each one reaches past the model's capability and grabs the relationship. Each one demands the same three things:
- Disclosure, on repeat. The product must tell the user, conspicuously and more than once, that it is an AI and not a person. China adds a break prompt after two continuous hours; California adds a recurring reminder for known minors.
- A self-harm off-ramp. The product must detect crisis signals and route them to real help — referral in the US statutes, active intervention in China.
- A different rule for minors. California and New York layer on extra reminders and protections; China bans companion services for minors outright and requires guardian consent under 14.
Three governments, working independently, drew the same line — and it runs across capability, not along it.
That convergence is the signal. When one jurisdiction invents a category, it's a policy experiment. When three do it without talking to each other, the category is real, and it is going to keep spreading. The US FTC already has a 6(b) inquiry open — orders went to seven firms, from OpenAI and Meta to Character Technologies and xAI — which is how a category graduates from state law to federal attention.
The load-bearing question is the definition#
Here is the part that matters for anyone building. The three laws agree on the obligations. Where they differ — and where your exposure actually lives — is the classification test: what counts as a companion in the first place.
China reaches services that "simulate a personality" to provide sustained emotional interaction. California scopes to "companion chatbots" that sustain an adaptive, human-like relationship. New York targets AI companions offering ongoing emotional or relationship interaction. None of them draw a bright technical boundary, because there isn't one. The difference between "a friendly assistant" and "a companion" is a matter of degree — persistence of persona, emotional register, whether the user comes back for the bond or the task.
That vagueness is not a loophole; it's the risk. It means the question "is my product a tool or a companion?" can no longer be answered by your design deck. It has to be answered against three separate statutory definitions, and it has to be answered before you cross a threshold — China's measures trigger a mandatory provincial assessment at 1 million registered users or 100,000 monthly actives, and the US laws bind you the moment you have users in-state.
What a founder should actually do#
You do not need a general counsel to get ahead of this. You need to run the classification test on your own product honestly, and if you land on the companion side of the line, build the three primitives in from the start rather than retrofitting them:
- Persistent, unmissable AI disclosure — at the start of a session and on a recurring cadence, not buried in a settings page.
- A crisis path — detection of self-harm language plus a referral to a real hotline, wired as a first-class response, not a content filter.
- A minors posture — at minimum age-aware reminders; at maximum, deciding you simply will not serve minors a companion product, which is now the only compliant answer in China.
The instinct in the industry has been to ship a "friend" while calling it an "assistant," because assistant sounds like a tool and tools were unregulated. That arbitrage is closing. If you build a bond, three governments have already decided the bond is the regulated surface — and the cheapest time to know which side of the line you're on is now, while you can still choose.
The models will keep getting better on both sides of that line. What changed between November and July is that the line got drawn three times, in three places, by people who never spoke to each other. That's not a coincidence you get to ignore.
Related: we turned China's version of the test into a self-audit you can run on your own product, and traced why both Chinese giants chose the off switch over compliance.
