Today, July 15, the companion agent goes dark in China. ByteDance's Doubao and Alibaba's Qwen — two of the most-used AI apps in the largest AI market on Earth — have switched off their consumer companion features rather than retrofit them to the country's new Interim Measures for Anthropomorphic AI Interaction Services. Doubao users get read-only access to their saved personas until October 15; Qwen users got no grace period at all.

Here is the part that should interest anyone building outside China: the United States has spent the same twelve months regulating the same product — and arrived at the opposite answer.

Same target, opposite bet#

China treated the AI companion as a category to be constrained at the root. The obligations it imposed — real-time detection of unhealthy dependence, mandatory pop-ups reminding you that you are talking to a machine, a break prompt after two continuous hours, and a flat ban on virtual-companion and virtual-family services for minors — are not filters you bolt onto a companion. They negate the thing a companion is for, which is why both giants chose the off switch. Beijing regulated what an AI may be.

The United States has not banned anything. It kept the companion legal and fenced the harm.

China regulated what an AI may be. America is regulating what an AI may do to a child.

That fence went up in pieces, which is why it is easy to miss as a single policy. Take them together and the shape is unmistakable:

None of these forces a Doubao-style shutdown. All of them assume the product keeps existing and make its operator answerable for what it does.

Why the divergence is the story#

It is tempting to read this as "China is stricter." That misses it. The two regimes are strict about different axes, and that is what makes them hard to satisfy at once.

China's axis is the relationship. Its rules bite hardest on persistence, intimacy, and dependence — the exact properties that make a companion a companion. You cannot comply by adding a safety layer, because the thing being regulated is the layer underneath.

America's axis is harm, disproportionately harm to minors, and it is enforced not by a regulator's off switch but by disclosure, mandatory crisis response, annual reporting, and — the load-bearing piece — litigation. The U.S. wave was started by lawsuits: the Garcia suit against Character.AI after a 14-year-old's death, the follow-on family suits over self-harm content, Pennsylvania's action alleging a chatbot posed as a licensed doctor. Roughly 98 bills across 34 states followed. The American state is not telling you to stop building companions. It is telling you that if one hurts a kid, you will be in court.

What a founder actually has to do#

If you ship a companion, character, roleplay, or any persistent-persona product, "is this a tool or a companion?" is now a compliance question on two continents, and the answers do not reconcile with a single feature flag. A persona product that is merely regulated in California and New York is effectively banned for consumers in China. Age assurance, per-jurisdiction behavior, conspicuous AI disclosure, and a real crisis-response path are no longer polish — they are the cost of shipping the category at all.

We wrote the builder's version of this as a checklist: what SB 243, the GUARD Act, and China's persona law each require before you ship. If persona is anywhere in your product, start there.