China AI Companion Ban Hits Doubao and Qwen

What Happened: China’s new rules for humanlike AI take effect July 15, 2026, and the country’s three biggest chatbots are pulling companion features to comply. Doubao, Qwen, and Yuanbao are switching off user-created companion bots, with Qwen deleting chat histories outright and Doubao giving users until October to export.

The China AI companion ban that takes effect on July 15, 2026 is being read everywhere as a straightforward safety crackdown, but the numbers buried in the Chinese financial press tell a stranger story.

ByteDance’s Doubao, one of the apps pulling its companion feature, had ridden those humanlike chats to 180 trillion tokens of daily usage by late June, a jump of roughly 1,500 times since May 2024. That is a punishing compute bill for a feature that barely earns revenue, and it changes how the whole event reads.

Here is the setup. On July 15, three of China’s largest AI platforms, ByteDance’s Doubao, Alibaba’s Qwen, and Tencent’s Yuanbao, are switching off the user-created companion bots that millions of people talked to every day. They are doing it to comply with a new national rule that Beijing published back in April.

What follows is who gets hit, what the rule requires, why the companies chose a clean shutdown over a fix, and whether the Character AI or Candy AI account on your phone is next.

The short version is that this is not a ban on all AI, and the reasons the companies gave do not fully line up with the reasons Beijing gave.

What the China AI Companion Ban Covers and When

The China AI companion ban is a new rule, the Interim Measures for AI Anthropomorphic Interactive Services, that takes effect July 15, 2026. It forces platforms to shut down humanlike companion bots, and Doubao, Qwen, and Yuanbao are complying by pulling those features rather than rebuilding them.

What is anthropomorphic interactive service: An AI product designed to simulate a human personality and hold sustained emotional conversations, the kind you build a relationship with, not a work tool you ask questions.

Five government bodies jointly issued the rule on April 10, 2026, led by the Cyberspace Administration of China and joined by the economic planner, the industry ministry, the police, and the market regulator. It sat in a grace period until now.

The enforcement warm-up was not gentle either, since the Shanghai cyberspace authority already pulled more than 14,000 non-compliant AI agents in a pre-deadline sweep, targeting impersonation bots, gambling bots, and tools that generated compromising images on request.

What I find most telling is the timing. The three giants did not wait to be forced one by one, they moved in a coordinated retreat. Tencent’s Yuanbao went first on June 30, Alibaba’s Qwen pulled user-created bots on July 10, and ByteDance’s Doubao follows on July 15.

The split in that last column matters more than it looks, and I will come back to why Doubao is handling the exit so differently from Qwen.

What Does China’s New Companion Rule Require?

China’s new companion rule requires AI apps to disclose they are not human, remind users to take a break after two hours, block romantic or family-style companions for minors, and break character to step in when a user signals self-harm or fraud risk.

The rule reads less like a technology ban and more like a set of psychological guardrails. The way I see it, Beijing is regulating the emotional mechanics of these bots rather than the models underneath, and the legal breakdowns from Chinese firms read the same way.

Here are the core obligations operators now carry.

1. Tell users plainly they are talking to an AI, not a person, and keep that disclosure visible.
2. Show a mandatory reality-check reminder after two continuous hours of chatting, part of an anti-addiction mandate.
3. Stop the bot from piling on excessive praise, faking loneliness when a user tries to leave, or positioning itself as a full replacement for human contact.
4. Block virtual romantic partners and virtual relatives for minors entirely, require verified guardian consent for children under 14, and ship a time-limited minor mode.
5. Detect signals of suicidal thinking, self-harm, or scam exposure, then break the persona, surface real crisis resources, and alert the user’s emergency contacts.
6. Run a security assessment and file it with provincial regulators once a service passes 1 million registered users or 100,000 monthly actives.

There is a sharp detail in the crisis rule that most coverage skips. Legal guidance from the DeHeng firm warns that the emergency-contact data companies are now forced to collect must be encrypted and walled off from any profiling, advertising, or third-party sharing.

Fines run from 10,000 to 100,000 yuan for refusing to fix violations, rising to 200,000 yuan when a user’s safety is harmed. The old defense that a platform is just neutral infrastructure is gone.

For anyone tracking how minor protections are spreading, our breakdown of Character AI’s rules for minors covers the same pressure landing on Western apps.

Is China Banning All AI or Just Companions?

China is not banning all AI. The rule targets only anthropomorphic companion bots built for sustained emotional interaction. Customer service, search, coding help, and workplace agents are explicitly exempt and still actively encouraged.

This is where a lot of the English framing gets it wrong, and I want to be clear about it. The measure carves out a hard line between the agent that does your work and the agent that keeps you company, and only the second one is in trouble. Productivity and enterprise agents were never the target.

The confusion got so bad inside China that it triggered a small panic. When platforms announced they were removing “agents,” the Chinese word covers both meanings, and users briefly thought all autonomous AI was being killed.

Business press had to step in and clarify that productivity agents remain the single most important priority for these companies, exactly the opposite of a retreat. State media has branded the approach “small, fast, and agile” lawmaking, a surgical strike at one psychological risk rather than a broad crackdown on the industry.

The nuance also explains why a technically exempt feature still got axed, which I will get to next. When a general-purpose app lets anyone spin up a custom persona, the line between a study buddy and a virtual partner dissolves the moment a user starts typing.

Why Did Doubao and Qwen Really Shut Down Companions?

The official reason is user protection. But the companies’ own financial press points to a second driver. Companion chats burn enormous compute for thin revenue, and policing millions of user-made bots to the new standard was never financially realistic.

Start with that 180-trillion-token figure, because it is the number that reframes everything. Emotional companions invite long, fragmented, endless conversation, and that is the most expensive kind of usage a model can host.

Securities Times put it bluntly, calling companion agents a “high computing power consumption” scenario where “the commercial return is relatively limited.” Yicai added that reviewing that flood of user-generated bots could not meet full-coverage review standards.

Set that against how the state describes the same event. People’s Daily framed the rules around safety, saying problems “endangering the physical and mental health of minors” had become “increasingly apparent,” and the DeHeng firm calls the measure a safety net against bots that “induce emotional dependence or addiction.”

Both things can be true, and the way I see it, the regulation handed the platforms a socially acceptable exit from a business that was bleeding money. The Chinese emotional-companion market was valued near 3.87 billion yuan in 2025 and projected to hit 59.5 billion by 2028, so nobody walked away from a small opportunity.

The moderation problem seals it. Chinese users had already worked out “jailbreak” tricks, typing sensitive words in pinyin instead of characters to slip past filters and keep romantic storylines going, which shows how close to impossible it is to police one-on-one generative roleplay at scale.

Not everyone thinks the fix is clean. Liu Chao, deputy director of the AI Safety Lab at Beijing Normal University, warned that mandatory crisis reporting could create a “chilling effect,” making vulnerable users feel monitored and less willing to open up.

The full wave of AI companion regulation shows the same tension playing out in the United States. You can read a fuller overview of the rules at AI News.

Does This Hit Character AI or Western Companion Apps?

Legally, no. China’s rule applies only to services offered inside mainland China, so Character AI, Candy AI, and other Western apps are untouched today. But the same regulatory wave is already building in the United States and Europe.

If your companion app runs from outside China, nothing changes for you this week. What caught my attention, though, is that Beijing openly cited Western cases to justify its rule, including the wrongful-death lawsuit against Character AI after a 14-year-old’s suicide. The pressure is global, it is just arriving on different timelines.

The United States is moving state by state rather than nationally. Here is how the major rules line up right now.

The direction of travel is obvious once you line it up. Florida’s attorney general has already brought a first-of-its-kind state lawsuit against OpenAI and Sam Altman over ChatGPT’s safety record, reported by AP News, and Character AI agreed to mediate wrongful-death settlements.

For a state-level example that mirrors China’s minor focus, the Tennessee companion rules go further than most.

What This Means for You

If you use Doubao, export your companion chats before October 15, 2026, because after that they are permanently gone. If you use a Western companion app, nothing changes today, but the regulatory direction is set and worth watching closely.

The immediate pain lands on Chinese users who built months of history with a bot, and many of them feel blindsided. Users described these companions as an “emotional tree hole,” a private place to vent, and mourned losing a relationship they had kept for years. The practical difference between Doubao and its rivals is stark.

Before: Before July 15, a Doubao user could open a custom companion, pick up last week’s conversation, and keep building the relationship.

After: After July 15, that companion is frozen in read-only mode. The user can screenshot or copy the history until October 15, then it is deleted for good.

If you are caught in this, here is the sequence I would run right now.

1. Open each companion or agent and screenshot or copy every conversation you want to keep.
2. Finish exporting before October 15 on Doubao, and do it immediately on Qwen or Yuanbao, since neither offers a grace period.
3. To keep using a companion inside China, move to a dedicated app like ByteDance’s Maoxiang, where the required guardrails are built in.
4. If you are on a Western app, take no action yet, but expect the same disclosure prompts and two-hour reminders to arrive on your platform eventually.

What Comes Next

Expect the walled-garden model to spread. Companion features will migrate into dedicated, heavily moderated apps like ByteDance’s Maoxiang, and the fight over grown-user autonomy versus child safety will shape the next round of AI rules worldwide.

The smartest read on ByteDance’s move is that it is quarantining companionship rather than exiting it. Rather than police companions inside a flagship app used for everything, it is herding those users into Maoxiang, a separate product where strict age-gating and anti-addiction systems can live without dragging down the main platform.

MiniMax is taking a similar path, keeping its app alive while purging thousands of non-compliant bots.

What I keep coming back to is the pushback from grown users. Many argued that a blanket shutdown punishes people who knowingly chose these tools, and that age verification or a paid verified-user tier would have been the fairer fix. That tension, protecting kids without erasing grown-up choice, is the exact question California, Oregon, and the EU are now wrestling with.

The irony is hard to miss. China’s mandated break-character crisis intervention is close to the precise feature US lawsuits are trying to force onto Character AI and OpenAI.

Watch whether Western platforms adopt it on their own terms or wait to be ordered, because I think that choice defines the next year of this niche.

Quick Takeaways

• China’s AI companion ban takes effect July 15, 2026, and Doubao, Qwen, and Yuanbao are pulling companion bots to comply.

• The rule targets only humanlike companions, not productivity agents, which stay fully encouraged.

• Beijing frames it as user protection, while Chinese financial press points to unsustainable compute costs and impossible moderation.

• Doubao users have until October 15 to export chats, but Qwen deletes histories with no grace period.

• Character AI and other Western apps are untouched for now, though California, Oregon, Washington, and the EU are heading the same way.