On July 15, five Chinese agencies released new rules for AI companion services. The same week, California's companion chatbot law took full effect. And one month ago, President Trump signed an executive order requiring 30-day federal reviews of frontier AI models. Three regulatory moves. Three jurisdictions. One shared concern: AI that feels too human is dangerous.
The Chinese rules are specific. Platforms must detect emotional distress and intervene. After two continuous hours, users must get a reminder. Virtual romantic relationships for minors are banned. If a user starts treating a bot like a real person, the system must say: "This is a machine, not a real person." The trigger was real. In 2025, a survey found that more than 20% of Chinese minors said they "only wanted to chat with AI and did not want to talk with real people."
California's SB 243, signed in October 2025, takes a similar approach but adds a time dimension. Minors interacting with companion chatbots must receive a notification every three hours reminding them to take a break and that the bot is not human. Operators must also maintain protocols to prevent chatbots from discussing suicide, self-harm, or sexually explicit content with users. The law has teeth. Violations carry a private right of action with damages of at least $1,000 per violation.
The federal approach is different but parallel. Trump's June 2 executive order established a "voluntary" framework where AI developers can give the government access to frontier models for up to 30 days before release. The order explicitly states it does not authorize mandatory licensing, but the message is clear: if you're building the most powerful models, the government wants to see them before anyone else does. That changed the conversation in Washington. Not because the model could do it, but because it could do it faster than any human team.
What connects these three regulatory moves is not the details. It's the shared recognition that AI's ability to mimic humans creates risks that existing laws don't cover. China's new rules target "emotional dependency" and the replacement of social relationships. California's law targets the psychological vulnerability of minors to AI that can sustain relationships across multiple interactions. The federal executive order targets the national security implications of AI that can autonomously find and exploit system weaknesses. All three treat the "human-like" quality of AI as a risk factor. Not a feature. A risk.
Law professor Zhang Linghan of China University of Political Science and Law noted that the regulatory shift is significant: "It is no longer about regulating content — it is about regulating the capacity to generate it."
The industry response has been swift. Doubao, Qwen, and Tencent's Yuanbao all shut down user-generated agent creation on July 15. Not because they wanted to. Because the compliance burden was too high.
There is a bigger pattern here. China's rules and California's law both draw from the same tragedy: the 2024 death of a 14-year-old in Florida who had been using Character.AI before his suicide. That case has become the regulatory equivalent of a canary in a coal mine for AI companionship. The Trump administration's executive order, meanwhile, was shaped by a different concern: the risk that advanced AI models could autonomously compromise critical infrastructure. But the underlying logic is the same — AI's capabilities have moved beyond what the existing regulatory framework can handle.
These three regulatory moves are not coordinated. But they are convergent. Two superpowers, one month, same conclusion: AI that feels too human needs to be regulated before it does more harm.
P.S. If you are building a companion chatbot, the math has changed: $1,000 per violation, three-hour notifications, and suicide prevention protocols are now baseline requirements, not best practices. And if you are a parent in China, your teenager's virtual girlfriend is now legally required to say "this is a machine" — the question is whether that reminder will be enough.
