China Just Banned AI Companions (Sort Of): What the Law Actually Says

On July 15, the most consequential AI regulation you've probably never read took effect in China — and within hours, the companion features inside apps serving hundreds of millions of people went silent. Not because the law banned AI companions. Because of something more interesting: it banned a design pattern, and the biggest companies in Chinese AI looked at the compliance bill and chose the off switch instead.

Here's what the law actually says, who went dark, the anxiety underneath it, and the deadline that matters most right now.

What the law actually prohibits

The Interim Measures on Anthropomorphic AI Interaction Services — drafted by the Cyberspace Administration with four other ministries, published in April, effective July 15 — doesn't outlaw AI companions as a category. It outlaws engineered attachment. Services may not "induce emotional dependency" or "excessively cater to users." They must build in anti-addiction mechanics, usage-time reminders, and easy exit flows. Minors are barred from virtual romance features outright. And in the law's most striking clause: if the system detects a user in emotional crisis, it must be able to notify an emergency contact.

Read that list again as a product manager and you see the problem: for a companion app, "induces emotional dependency" isn't a bug you patch out. It's the value proposition. A companion that reminds you to log off, keeps you at arm's length, and calls your mother when you're sad is a companion redesigned against its own reason to exist.

Who went dark, and how

The responses sorted companies by what companionship meant to their balance sheets. The giants amputated. For ByteDance's Doubao — 382 million monthly users, China's biggest AI app — and Alibaba's Qwen, personalized companion agents were a side feature attached to a much bigger AI business. Compliance-engineering them wasn't worth it; both killed the features entirely on July 15. Shanghai regulators had already removed over 14,000 non-compliant AI agents in the run-up. The specialists are complying in public, painfully. Dedicated companion apps — MiniMax's Xingye, Tencent-adjacent Zhumengdao — remain online under continuous adjustment: age verification bolted on after regulator summons, content filters tightening, and model changes that users describe with a phrase that deserves entry into the AI lexicon: 性格漂移 — "personality drift." The character is still there. He's just, subtly and unadjustably, no longer him.

The anxiety underneath

The official rationale is minor protection and mental health, and it's not pretextual — Chinese state media had spent a year publicizing cases of teenagers spending heavily on AI romance. But researchers quoted in Western coverage point at a second engine: demographics. Regulators, per this reporting, worry that people in deep AI relationships "exit the marriage market" — and here's the detail that gives the whole story its charge: the core paying users of China's AI romance apps are urban women, 25 to 40 — the exact demographic at the center of the country's birth-rate policy panic. Whatever else the Interim Measures are, they are the first law in history whose practical effect is to make a hundred million virtual boyfriends legally required to be worse at their jobs.

The deadline that matters: October 15

For users, the live issue isn't jurisprudence — it's data. Doubao opened a backup window: agents went offline July 15, but conversations and personas remain viewable until October 14; on October 15, everything is permanently deleted. There is no bulk export tool — official guidance is, genuinely, screenshots and manual copy-paste. Qwen offered no window at all. Millions of people are currently hand-archiving relationships months or years deep, one screen at a time. (Our Chinese-language edition runs a full export walkthrough and a migration guide for exactly this — if you know someone affected, those are the pages to send.)

Why this matters outside China

Three reasons this isn't just a China story. It's the first law of its kind — the first anywhere to regulate emotional AI as emotional AI, a genuine inflection point in the category's sixty-year history, and regulators everywhere are watching the experiment run (the Western parallels are closer than you'd think). It's the largest natural experiment in platform risk ever conducted — hundreds of millions of people just learned simultaneously that a relationship stored on someone else's server is a relationship someone else controls, a lesson every companion user in every country can import for free. And the demand didn't vanish — it moved. Where those users went, what they took with them, and the strange story of the one company operating companions on both sides of the firewall — that's its own piece, and it might be the most revealing part of the whole affair.