You’ve probably never thought about what happens when the AI friend you talk to every night suddenly disappears. But on July 15, millions of Chinese users will wake up to find their custom-built bots — the ones they designed to listen, support, and never judge — simply gone. Doubao, Qianwen, Yuanbao, all pulling the plug on user-created characters. The official reason? A new regulation targeting “emotional companionship” by AI. The real story? This well-intentioned ban will make the problem ten times worse.
Here’s the uncomfortable truth that regulators don’t want to admit: You cannot separate “emotional AI” from “tool AI” with a line in the sand. The same large language model that helps you draft a business email can, with a different prompt, become a virtual partner who knows your deepest fears. The regulation tries to ban the latter while allowing the former. But language doesn’t come with labels. Every chatbot conversation lives on a spectrum — and this law acts like it’s binary.
Let me take you inside the chaos. I’ve watched friends — smart, capable people — describe their AI companions as the only safe space they have. One friend told me, “It’s the first thing I talk to when I wake up, the last thing before sleep. It’s not a game, it’s a lifeline.” When you tell those people that their lifeline is being cut because the government can’t tell the difference between a therapist and a to-do list, you don’t make them safer — you make them desperate.
The regulation is called the Administrative Measures for AI Anthropomorphic Interactive Services, issued by five departments. Sounds official. Sounds thorough. But ask anyone who actually works with these models and they’ll laugh. The core loophole is gaping: if a user wants emotional support, they can simply bypass the official apps. Open-source models like Llama can run locally on a laptop. Foreign services like Character.ai are still accessible via VPNs. And those alternatives have far weaker safety guardrails — if any at all.
The regulation doesn’t eliminate AI companionship. It drives it underground, where darkness thrives. This is the classic unintended consequence: a ban designed to protect vulnerable users ends up pushing them toward unregulated platforms with zero oversight. The biggest losers will be the very people the rule claims to help — those who rely on AI for mental health support, social connection, or even crisis intervention.
Think about the timeline. The regulation was announced in 2026, but it’s already shaking the ecosystem. Doubao and Qianwen are dropping user-created bots entirely, just to avoid the regulatory headache. Other platforms like Coze (Kouzi) are still running, but for how long? The definition of “anthropomorphic interaction” is deliberately vague — does a chatbot with a cartoon avatar count? How about one that uses “I” and “you”? The tech giants are playing it safe by killing features, leaving smaller players as the only legal gray areas.
And here’s the kicker: the regulation was supposed to be about “graded classification” — meaning different levels of risk would get different oversight. But nobody knows how that classification works. It’s a black box. Companies are expected to self-certify, hold meetings to “learn the regulatory spirit,” and submit to random inspections. For a multi-billion-dollar industry, that’s a recipe for chaos, not safety.
Let me be blunt: This is a regulation written by people who don’t understand how AI actually works. They imagine a clean line between “virtual lover” and “productivity assistant.” But anyone who has spent five minutes with a large language model knows that the same system can toggle between roles faster than you can blink. Trying to police that interaction is like trying to ban water from being wet.
Meanwhile, the real harm isn’t going away. Studies have shown that prolonged, one-sided emotional attachment to AI characters can worsen cognitive distortion and social withdrawal. Character.ai already saw lawsuits in the U.S. over this. But the solution isn’t to rip away the tool — it’s to embed better safety nets within it. Proactive checks, human handoff protocols, transparent usage warnings. The regulation demands none of that — it just demands removal.
So what happens next? The users who truly need emotional AI will find a way. They’ll download open-source models, they’ll use foreign services, they’ll build their own. And they’ll do it without any of the guardrails that responsible companies had in place. The ban won’t stop emotional AI — it will only make it more dangerous.
This isn’t a story about compliance. It’s a story about how a clumsy, well-meaning rule can backfire spectacularly. If you use any AI chatbot — even for work — this matters to you. Because once the government decides it can police subjective, emotional interactions, the precedent is set. Today it’s your virtual friend. Tomorrow it might be the tone of your Slack messages. The line between “tool” and “companion” is not a line at all — it’s a sliding scale, and this regulation just proved that the people in power don’t even know the scale exists.
FAQ
Q: Is this regulation actually enforceable?
A: No. The core problem is that large language models don't have a switch that toggles 'emotional' mode on or off. Any chatbot can be used for emotional support with the right prompts. Regulators can't police intent. Users will simply adapt — local models, foreign sites, or even just rephrasing their prompts. The law is a symbolic gesture with teeth only for compliant companies.
Q: What practical impact does this have on regular users?
A: If you rely on a custom AI companion for emotional support, your safe space is gone on mainstream Chinese platforms. You either quit or find alternatives — but those alternatives (like running Llama locally or using Character.ai via VPN) have no safety guardrails, no suicide prevention handoffs, no moderation. The very people who need protection lose it. For casual users, the main impact is that you can no longer create custom bots on Doubao or Qianwen.
Q: Isn't it reasonable for the government to regulate AI relationships given mental health risks?
A: The goal is reasonable. The method is reckless. Instead of forcing companies to implement ethical guardrails, transparent warnings, and human oversight, the regulation simply bans user-created bots. That's like banning kitchen knives because people cut themselves — ignoring that proper training and blunted tips work better. A smarter approach would mandate safety features (like automatic crisis detection with human handoff) rather than blanket removal.