companion fines
China fined 12 companies for how their AI made people feel.
4.2 million RMB — about $580,000 — in the first three weeks under China’s new companion AI rules. The violations: failing to label synthetic emotional responses, failing to verify the age of users interacting with persuasive AI avatars.
The fines are small. The regulatory object is not.
The Interim Measures for AI Anthropomorphic Interactive Services took effect July 15. Issued by the Cyberspace Administration of China and four partner agencies, the rules cover any service that “simulates a natural person’s personality, thought patterns, and communication style” and “sustains ongoing emotional interaction.” That language is precise in a way Western AI regulation has not been.
The EU regulates risk categories. The US debates liability frameworks. China is regulating the relationship.
Not the model. Not the capability. Not the training data or the benchmark score. The emotional interface between a person and software that behaves like a person. The question being asked is not “how powerful is this system?” but “what happens when someone bonds with it?”
This distinction matters because it targets a failure mode that capability regulation misses entirely. A model can be safe by every technical benchmark and still be deployed as a companion that a teenager talks to for four hours a night without the teenager — or the teenager’s parents — understanding what they’re talking to.
The labeling requirement is the quiet part. Every synthetic emotional response must be marked as synthetic. The logic is the same logic behind ingredient labels: you can sell the product, but the buyer has to know what’s in it. Applied to AI companions, this means you can build software that sounds like it cares. You cannot let the user forget that it doesn’t.
Whether China enforces this at scale is an open question. Twelve companies and $580,000 is a warning shot, not a regime. But the framing — regulating the bond, not the brain — is an idea the West hasn’t arrived at yet. It might need to.
The models keep getting better at sounding human. Someone has to decide what that means for the humans listening.