Headlines this weekend screamed "China bans AI." That's wrong — and the truth is more interesting.

China isn't banning AI. It isn't even banning AI agents.

It's banning one specific thing: the agent that keeps you company.

On July 15, a new rulebook takes effect targeting AI that simulates a human personality for sustained emotional interaction — AI girlfriends, AI companions, AI "family members," persistent role-play personas. The kind you talk to every night.

And rather than rebuild those products to comply, ByteDance and Alibaba are just… turning them off.

Millions of users. Custom companions built over months. Gone in days.

Let's name names.

ByteDance's Doubao — one of China's biggest consumer AI apps — told users Friday night its agent feature goes offline July 15, blaming "product function adjustments." You get read-only access to your companions and chat histories until October 15. After that? Unrecoverable.

Alibaba's Qwen went harder. Humanlike and user-created agents die July 10. Broader agent functions July 15. And here's the brutal part: permanent deletion of your configurations and conversations, with no migration path announced. No export. No fallback. Gone.

Tencent's Yuanbao already quietly pulled a similar feature back in June.

So three of China's tech giants, same move, same window.

And on Weibo, people are grieving. One user called their agent long-standing emotional support and asked why there's no easy way to save the chat histories.

For some people, these weren't features. They were relationships.

The law is the Interim Measures for the Administration of AI Anthropomorphic Interaction Services.

It was co-issued back in April by China's internet regulator (the CAC) plus four other agencies. It takes effect July 15.

Here's what it demands of companion AI:

Read that list again. Every requirement fights against the product.

This is the part most coverage misses — and the part you'll actually want to think about.

Doubao and Qwen didn't break a ban. They hit a design conflict.

A companion agent is built to remember you, stay consistent across sessions, and keep an ongoing relationship going. That's the entire value proposition.

Now try bolting on mandatory anti-addiction interrupts, constant AI-disclosure reminders, and real-time dependence detection.

You can't. Persistent-memory attachment and "we must interrupt you if you get too attached" are opposite goals.

So the companies made a call: don't retrofit. Rip it out.

That's not a regulation story. That's an architecture story wearing a regulation story's clothes.

Before you think "China killed AI agents" — don't. Beijing drew a very precise line.

Customer-service bots? Exempt. Knowledge Q&A? Exempt. Workplace assistants, education tools, research tools? All exempt — as long as they don't drift into ongoing emotional attachment.

The enterprise agent economy isn't just spared. It's being actively encouraged through national standards work.

In China's agent economy, the companions go first and the workers stay.

And enforcement is already live. On June 26, Shanghai authorities yanked more than 14,000 non-compliant agents tied to impersonation and privacy violations — a warning shot two weeks before the deadline.

One more thing worth knowing: this isn't a uniquely Chinese crackdown. California's SB 243, the first US companion-AI law, took effect January 1. The difference is speed and teeth — Beijing is enforcing hard and mandating that the data actually gets deleted.

Two governments, same fear: software built to make you attached.

China's Interim Measures for the Administration of AI Anthropomorphic Interaction Services take effect July 15, and instead of complying, ByteDance's Doubao and Alibaba's Qwen are shutting down their companion-agent features entirely — Doubao going read-only until an October 15 data wipe, Qwen permanently deleting user data with no migration path, and Tencent's Yuanbao having pulled a similar feature in June. The rules target AI that simulates human personality for sustained emotional interaction (companions, AI partners, virtual family), while explicitly sparing customer service, workplace, education, and research agents. The reason the giants killed the feature rather than fix it: mandatory anti-addiction systems, AI-disclosure notifications, and dependence-detection are fundamentally incompatible with agents built on persistent memory and ongoing relationships. This is confirmed, not rumor — South China Morning Post broke it, China's state-run Global Times confirmed the regulatory motive, and TechNode and TNW have detailed accounts. The through-line: China isn't banning AI, it's banning the AI that keeps you company — a surgical move on parasocial products, landing the same month California's own companion-AI law took effect. Next to watch: whether compliant companion apps like ByteDance's Maoxiang bring these products back in a safer form, and how hard enforcement bites after July 15.

Stay building. 🤖