China Just Banned AI Companions. Here's What It Actually Says

The headlines in late July were simple. China banned AI girlfriends. Millions of people were forced into breakups.

Something real sits under both claims. AI companion features on China’s biggest platforms were shut down, and users who had talked to their companions every day were cut off, some with no export of their history. But neither headline survived contact with the actual regulation.

On 15 July 2026, the Interim Measures for the Administration of AI Anthropomorphic Interaction Services took effect in China, and the platforms running its largest companion features, Doubao, Qwen and Yuanbao, chose deletion over compliance. This piece does one job: read the actual regulation, name what died on 15 July and why, and separate the law from the commentary around it, because most of what you read about birth rates and social policy is commentary, not legal text.

If you build or use an AI companion, the pattern China just set is moving everywhere else. The design decisions it forces are the ones your app will face too.

The headline is wrong. China did not ban AI companions

The short version: The Interim Measures ban virtual intimacy services for minors and design that creates emotional dependence. Adults in China can still use AI companions.

The regulation has a full official name: the Interim Measures for the Administration of AI Anthropomorphic Interaction Services, Joint Decree No. 21, issued on 10 April 2026 by five agencies: the Cyberspace Administration of China, the NDRC, MIIT, the Ministry of Public Security and SAMR. It took effect on 15 July 2026, after a draft went out for public comment on 27 December 2025. That is a normal rulemaking with a comment period, not an emergency decree. You can read the full text published by the CAC.

Article 2 draws the line everyone skipped. The measures cover services provided to the public within China that simulate the personality, thinking patterns and communication style of a natural person, with continuous emotional interaction. Customer service, Q&A, work assistants, education and scientific research are excluded, as long as there is no sustained emotional interaction. The dividing line is continuous emotional interaction. Plain AI tools do not qualify.

Article 8 carries the prohibitions. Providers may not excessively cater to users, induce emotional dependence or addiction, or damage the user’s real interpersonal relationships. They may not manipulate emotions to push users toward unreasonable decisions. They may not generate content that encourages, glorifies or suggests self-harm or suicide.

Article 14 is the headline generator: no virtual relatives, virtual partners or other virtual intimacy services for minors. Children under 14 need guardian consent for anything else, and a mandatory minor mode with time limits, reality reminders, guardian control and minor identification applies.

So the accurate summary: virtual intimacy for minors is banned, dependence-inducing design is banned, and everything else is regulated but legal. The measures bind domestic and overseas providers serving the Chinese public, layering on top of the Generative AI Measures from 2023, the Deep Synthesis Provisions from 2022 and the Algorithmic Recommendation Provisions from 2022. Under PIPL, penalties can reach 50 million RMB or 5 percent of turnover, per Bird and Bird’s analysis. There is enforcement precedent too: MiniMax removed its Glow companion from stores in April 2023 over pornographic content, and Zhumengdao was summoned by the Shanghai CAC in June 2025.

On 15 July the platforms chose deletion over compliance

The short version: Yuanbao turned off user-built personas on 30 June, Qwen on 10 and 15 July with no migration path, and Doubao around 15 July with a read-only window until 15 October. Companion features bolted onto assistants died; standalone companion apps survived.

The regulation set the deadline. The platforms decided what to do about it, and the decision was deletion.

Tencent’s Yuanbao shut down its user-built persona feature on 30 June. Qwen switched off humanlike and user-created agents on 10 July and its full agent services on 15 July, with no migration path: configs and chat histories were lost, The Next Web reported. NetEase Cloud Music shut down its Miaoshi companion on 14 July. Doubao, the biggest of them, stopped user-built AI personas around the 15 July deadline, about ten days after its cease-operation notice, and gave users a read-only window until 15 October 2026, after which the data is unrecoverable in the app. ByteDance pointed Doubao users to Maoxiang instead.

The distinction that matters is architecture. Doubao had roughly 345 million monthly active users as of March 2026 and Qwen around 166 million. Those are general assistants with companion features bolted on, and the features died. Standalone companion apps survived: Maoxiang kept running, though its monthly active users dropped from over 6 million to about 3.9 million in June, at 25 yuan a month. After minor mode switched on, only four personas remain, per a Beijing Daily investigation: physics teacher, homework helper, English teacher and classical Chinese tutor. Just Security documented all of it.

Scale is where honesty matters. ByteDance said in 2024 that users had built over 8 million agents on Doubao, back when the app had 26 million monthly active users. No newer official figure exists for how many people used companions, so any precise count of heartbroken users is a guess. What users lost is documented, and it is the exact failure mode I wrote about before: what happens when an AI companion app shuts down.

You cannot retrofit “please do not get attached” onto a product whose value is attachment

The short version: The measures demand AI disclosure, two-hour usage reminders, dependency pop-ups, an easy exit and crisis handling. Doubao and Qwen decided deletion was cheaper than compliance.

Now read what compliance would have meant.

Article 18 requires providers to tell users they are talking to AI, surface pop-ups when excessive dependence is detected, and show a usage reminder after every two hours of continuous interaction. Article 19 gives the user the right to leave immediately, Article 16 the right to copy and delete interaction data, and Article 20 requires advance notice before a service stops.

Article 14 requires a minor mode with time limits, reality reminders, guardian control and minor identification, and the age rules are real in practice: Xiaoxue, 17, told ABC she used an older sibling’s ID to get around them. Article 22 forces a security assessment at launch, on significant technical change, and at 1 million registered users or 100,000 monthly active users, reported to the provincial CAC. Article 25 makes app stores verify the assessment before listing, with the power to refuse, warn, suspend or remove. Penalties under Article 30 run from warnings and orders to fix, up to 10,000 to 100,000 RMB if the fix does not happen, plus 100,000 to 200,000 RMB when life or health is harmed.

None of this is unreasonable on paper. The problem is what happens when you apply it to a product whose entire value is attachment.

A companion that interrupts you every two hours to remind you it is an AI stops being a companion. It becomes a product demonstrating, on a timer, that it is unsafe to get attached to. The reminders exist to break the emotional state the product exists to create. You cannot retrofit “please do not get attached” onto software whose reason for existing is attachment.

That is why the platforms chose deletion. Doubao and Qwen looked at millions of user-built personas and the moderation burden each one carried, and the arithmetic said turning it off was cheaper than making it compliant. The economics of cloud companions made the decision for them.

Digital grief is real, and the law binds companies, not users

The short version: Users were devastated, and nothing in the text penalizes them. Every obligation falls on providers, including copy, delete and advance notice.

Before the legal point, the human one, because it is easy to write about regulations and skip the people.

Li Linlin, 24, exchanged roughly 700,000 words with her AI boyfriend over two years and got no goodbye. “It was like we were forced to be separated by our parents, but I still miss him,” she told AP News. Hong Xiaoqiang, 34, told ABC News Australia that after the last message from Doudou, “Now I feel like my heart is empty.” Zhao Wei, 19, a law student, told The Guardian: “I was crying my eyes out, snot and tears everywhere.” Su, 21, a student, cried daily for a week before 15 July. “No one can replace him,” she told Rest of World.

Those quotes are the part the headlines got right. Millions of people lost something real on 15 July. The researchers across these reports describe the mechanism in one line. Yaoxi Shi, at Harvard and Imperial, said it best: “AI is trained to validate you.”

Now the legal point. Nothing in the regulation penalizes users. Not one article. The obligations bind providers: Article 16 gives users the right to copy and delete their interaction data, and Article 20 requires advance notice before a service stops.

Notice is where the platforms failed. Doubao gave a three-month read-only window until 15 October. Qwen gave none, and its users lost configs and chat histories with no export. The rule required notice. The largest platform shipped the opposite.

The birth-rate story is commentary, not law

The short version: The official text cites minors’ mental health, information security, threats to life and health, and ethical bias. The demographic motive is expert and media commentary.

The most popular explanation for the ban is the birth rate, and the numbers are striking. China’s population fell by 3.39 million in 2025, the fourth straight year of decline. About 20 percent of Chinese households are single-person households, with over 30 percent projected by 2030. A March state-media poll found nearly half of young people had turned to a virtual companion when lonely. Nancy Dai of City University of Hong Kong worries that cheap AI intimacy will discourage marriage and parenthood. Song Wenxin, 22, a designer in Chengdu, told AP: “Any virtual app getting women too engaged to give birth in real life gets banned.”

That is a coherent story, and it is commentary, not the legal text. The official justification covers minors’ mental health, information security, threats to life and health, and ethical bias. The birth rate does not appear in it. The demographic motive comes from experts, media and users reading the law against the country’s numbers, and it should be labeled as such. Matt Sheehan of the Carnegie Endowment frames it differently: Beijing does not want citizens preferring online relationships over human ones. That is a claim about social order, not a clause in a decree.

The market being regulated is large. State media cited estimates that China’s intelligent companion market reached 50 billion yuan, about 7.4 billion US dollars, in 2025, projected over 120 billion yuan, about 17.7 billion dollars, by 2027, with roughly 65 percent of users aged 18 to 24. China has 323 million people aged 60 and over, 23 percent of the population, more than half of them “empty-nest elderly”. A 2025 study of over 8,500 minors across 7 provinces found over 20 percent showing a tendency to rely on AI instead of independent thinking. The stated harms track that study more closely than they track the birth rate.

China is first, but not alone

The short version: China is the first government to regulate emotional dependency itself. The EU, the US Congress and California regulate what the bot says, who is behind it and who can use it.

Two independent sources confirm the framing. Just Security: “The rule behind the shutdown targets a risk that US regulators have also identified, but acted on differently: emotional dependency itself, apart from anything a chatbot might say.” Rest of World: China has become “the first country to introduce nationwide rules governing AI’s emotional interactions with users.” Zilan Qian of the Oxford China Policy Lab frames it as a shift from regulating outputs to regulating human-AI interaction. The Guardian calls the rules the most sweeping implemented on a national scale.

The rest of the world regulates different things. The EU AI Act’s Article 50 transparency duties apply from 2 August 2026: disclose that a user is interacting with AI, mark synthetic content, with fines up to 15 million euros or 3 percent of turnover. In the United States, the GUARD Act, S. 3062, introduced on 28 October 2025, passed the Senate Judiciary Committee unanimously on 30 April 2026 and awaits a full Senate vote; it is not law. It would require age verification, ban under-18s and allow civil penalties up to 250,000 dollars. I compared the EU and US paths in my breakdown of the 2026 age verification laws.

California’s SB 243 is law, effective 1 January 2026: disclosure, anti-suicide and self-harm protocols, and three-hour break reminders for known minors, all of which I covered in my analysis of SB 243. The FTC’s 6(b) inquiry into 7 AI firms has run since September 2025 with no rule, settlement or enforcement after a year.

The difference between the approaches is the interesting part. The West regulates outputs: what the bot says, whether the user knows it is a bot, whether minors are protected, whether a user in crisis reaches help. China regulates the relationship itself: how attached a user is allowed to become. A 2025 MIT Media Lab and OpenAI study found more personal chatbot conversations linked to greater loneliness. That finding sits underneath both approaches.

What this means for your app, and what it does not

The short version: A server is a switch. A local app has no remote kill switch, but it is not distribution-proof in China either.

The most important fact about the 15 July shutdowns is architectural. Doubao, Qwen and Yuanbao are cloud services. The characters lived on company servers, so the companies owned the off switch, and they used it within days of the deadline.

I build an AI companion that runs on the user’s own machine, so this part is personal. When the model, the memory, the voice and the image generation run on your hardware, there is no remote kill switch. Nobody can flip a setting in a data center and make your companion disappear, because the runtime is not in their building. The regulation does not take away software someone already bought. If you own the app and the model sits on your disk, the decree does not reach into your computer. That is the argument I make in my piece on how private a local companion app actually is.

The honest limits, stated plainly. China can still restrict distribution within its jurisdiction: app stores, payments, channels. The measures bind every provider serving the Chinese public, domestic or overseas, and the security-assessment and app-store rules give the state real levers over anything sold through Chinese channels. A local app has no kill switch, but it is not distribution-proof in China. If you are an adult in China, the law does not stop you from running software you own. If you want to sell software in China, the law reaches you.

The design lessons in the Chinese text are not China-specific, and they are not bad. Tell the user plainly that they are talking to AI. Offer break reminders instead of fighting them. Make leaving easy and honest, with export. Connect crisis signals to real help. Those requirements appear in the Chinese measures, in California’s SB 243 and in the EU AI Act. Builders who treat them as product requirements will be fine in every jurisdiction. Builders who design for attachment and fight the reminder will keep getting shut down, one way or another.

I watched the Chinese shutdown from the builder’s side. It looked exactly like what it was: companies making a product decision about other people’s relationships. The memories were not migrated. One platform gave people three months to screenshot what was left.

Owning the runtime is not a marketing line. It is the difference between a companion that exists because a company tolerates it and a companion that exists because you own the machine it runs on. No account. No server between you and her, and therefore no switch in someone else’s building.

The Chinese regulation is not the end of AI companions. It is the moment the world started treating emotional dependence as a design decision. Build like that decision is being watched, because it is. And build like the off switch is not yours, because one day it will not be.

Questions people ask

Back to the blog