China Didn't Ban AI Agents. It Banned the Ones That Remember You.

 

Headlines this week made it sound like China was cracking down on AI agents wholesale. It wasn't. ByteDance's Doubao and Alibaba's Qwen both quietly shut off their companion-style agent features ahead of new rules taking effect July 15, but the assistants that do your work are untouched. It's specifically the agents built to keep you company that ran into trouble.

The distinction matters, and it's worth understanding why two of China's biggest consumer AI apps chose to kill a feature entirely rather than fix it.

China's new Interim Measures for the Administration of AI Anthropomorphic Interactive Services target services designed to simulate personality and sustain ongoing emotional interaction, the kind of agent that remembers you, stays consistent across sessions, and starts to feel like an actual relationship. Customer service bots, workplace tools, and education assistants are explicitly carved out, as long as they don't cross into sustained emotional engagement.

Here's the design conflict Doubao and Qwen ran into: the rules require anti-addiction systems, mandatory usage notifications, instant-exit mechanisms, and real-time detection of unhealthy dependence. Those requirements sit awkwardly with a product whose entire value proposition is making you want to come back and keep talking to the same persistent character. Rather than retrofit that tension, both companies just switched the feature off. Tencent's Yuanbao had already pulled something similar back in June.

Real people are absorbing the cost of that decision. Users took to Weibo describing these agents as long-term emotional support and mourning the sudden loss, with no easy way to export their chat history. Doubao is giving users a read-only window until October 15 before data becomes unrecoverable. Qwen users got no such grace period, agent data is simply set for permanent deletion.

The substance of the regulation itself is more thoughtful than "clampdown" suggests. Companion services are barred from serving minors without guardian consent, required to build usage-limited "minor modes," and mandated to detect signs of self-harm, suicidal ideation, or serious financial distress and escalate to guardians or emergency contacts. Engineering emotional dependence or using manipulation to push users toward decisions is explicitly prohibited. On the safety side alone, this is arguably a more complete framework than what currently exists in the EU, the US FTC, or California's SB 243.

But the rules leave real gaps too. There's no fixed technical definition of what counts as "emotional interaction," which is exactly why platforms chose to kill entire features rather than risk landing on the wrong side of an undefined line. The safety provisions are also bundled together with content-control and national-security requirements that answer to the state rather than the user, and there's still no clarity on how liability splits between platforms and the AI model providers behind them, or any right for users to actually take their own data with them.

So which half of this framework matters more? The safety half addresses harms that are already well documented elsewhere, including lawsuits over Character.AI's psychological impact on teenagers and regulatory scrutiny of Replika in Europe. The control half hands the state real leverage over what these systems are allowed to say, dressed in the same protective language.

Both halves are genuinely real. Which one a government is comfortable importing is the actual question every regulator watching this experiment will eventually have to answer.

Where do you land, does the safety framework here outweigh the state-control concerns, or is it the other way around?