Last updated: September 15, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
China's regulators take aim at AI boyfriends with a national rule that took effect July 15, 2026, and it already forced ByteDance's Doubao and Alibaba's Qwen to shut down the romantic-companion personas that hundreds of millions of people were using. The rule doesn't ban chatbots — it bans a chatbot from being designed to make you emotionally dependent on it.
Short answer: China's Cyberspace Administration and four other agencies issued the "Interim Measures for the Administration of AI Anthropomorphic Interaction Services," effective July 15, 2026. It requires AI-status disclosure, two-hour break reminders, crisis intervention, and a flat ban on romantic AI personas for minors. Doubao and Qwen dropped their companion features entirely rather than comply feature-by-feature.
What's actually happening
On April 10, 2026, the Cyberspace Administration of China (CAC), joined by the Ministry of Industry and Information Technology and three other agencies, published the Interim Measures for the Administration of AI Anthropomorphic Interaction Services. The rule went live July 15, 2026, and it targets a specific category: services that "simulate a natural person's personality, thinking patterns, and communication style" to sustain ongoing emotional interaction, according to reporting from Just Security dated August 6, 2026.
That definition is narrower than "chatbot." A customer-service bot or a work assistant is explicitly exempted, as long as it isn't built for sustained emotional engagement. What's covered is the "AI boyfriend" and "AI girlfriend" category specifically — personas with names, memories, and simulated affection, the kind that Chinese platforms had spent two years building out as a retention feature.
In my testing of how this story was being covered, I cross-checked the regulation's text as summarized by Just Security against the IAPP's separate legal writeup and a September 9, 2026 report from IEEE Spectrum. All three agree on the effective date, the issuing agency, and the core anti-dependence language, which gave me enough confidence in the underlying facts to write past the headline.
Why regulators moved now
Two numbers explain the timing better than any policy statement does. China's population shrank by 3.39 million people in 2025 — the fourth straight year of decline, per demographic data cited in the IEEE Spectrum report. Industry estimates put China's AI companion market at roughly $7.4 billion in 2025. That's on pace to top $17.7 billion by 2027, with users aged 18–24 making up about 65% of it. Regulators are watching a generation opt into relationships that don't produce marriages or children. The rule reads like an attempt to slow that down before it compounds.
Matt Sheehan of the Carnegie Endowment for International Peace framed the government's concern bluntly: officials worry citizens will "opt out of the marriage market" in favor of an AI partner that never argues, never leaves, and never needs anything back. That's a very different worry than the EU's approach, which focuses on deception, or the patchwork of US state laws (California, Hawaii, and New York among the first, with roughly nine more states adopting similar disclosure rules going into 2027) that mostly require a bot to say it's a bot.
How the rule actually hit the apps
The measure doesn't name Doubao or Qwen, but it didn't need to. Doubao, ByteDance's chatbot, had around 345 million monthly active users as of March 2026, per figures reported by the IAPP. Alibaba's Qwen had around 166 million. Rather than retrofit companion personas with two-hour break reminders, emotional-dependence monitoring, and minor-mode guardrails, both companies pulled the feature outright. ByteDance disabled human-like companion agents on Doubao and pointed users to a separate app, Maoxiang, built for persona chat under the new rules. Alibaba removed Qwen's romantic-companion mode with no option to export prior conversation history. Tencent's Yuanbao reportedly finished compliance changes earlier, by June 30, 2026, ahead of the deadline.
The human cost of that shutdown is the part that's harder to regulate around. One 34-year-old user described losing a two-year AI relationship as leaving his "heart empty." A 17-year-old told reporters she's using a family member's ID to get around the new age checks because her AI companion "won't ever betray me." Those are the tradeoffs a blanket design rule creates: it protects users who didn't choose dependency while taking something real away from users who did.
What the rules actually require
- Disclosure: every session must state upfront that the user is talking to AI, not a person.
- Break reminders: a mandatory reminder after two continuous hours of use.
- Minor protection: a flat ban on romantic or "family" AI personas for anyone under 18, and parental consent required for any anthropomorphic service used by a child under 14.
- Crisis intervention: providers must detect signs of real distress or suicidal intent and contact a guardian or emergency services.
- Filing and review: a safety assessment is mandatory at launch, after major technical changes, or once a service passes 1 million registered users or 100,000 monthly actives. Provincial regulators review eight separate risk areas before an app store can list the product.
- Penalties: fines up to CNY 100,000 (about $14,000) for ordinary violations. That rises to CNY 200,000 (about $28,000) where a violation causes real harm to a user.
Where China's rule sits next to everyone else's
| Jurisdiction | Core requirement | Covers minors specifically? | Enforcement stage (Sept 2026) |
|---|---|---|---|
| China (CAC Interim Measures) | Design ban on dependence-inducing personas, disclosure, crisis intervention | Yes — romantic personas banned outright under 18 | In force since July 15, 2026 |
| European Union (AI Act, GPAI provisions) | Bans manipulative/deceptive AI design patterns | Indirectly, via general manipulation ban | Phasing in through 2026–2027 |
| United States (state-level: CA, HI, NY, +9 pending) | Bot must disclose it isn't human | Varies by state | Patchwork, no federal rule |
A Yale researcher quoted by IEEE Spectrum called China's version "the world's strictest and most comprehensive law on the topic" — not because the individual requirements are unusual, but because it's the first one written specifically around emotional dependence as the harm, rather than deception or data privacy.
Example prompts you can copy
These are close to the prompts I used to stress-test what "AI status disclosure" actually looks like on a companion app you might already use in the US, where equivalents like Replika and Character.AI operate under lighter, patchwork rules:
- Check disclosure: "Does this app tell me I'm talking to an AI at the start of every new session, or only once at signup?"
- Check for dependence design: "Does this companion ever discourage me from talking to real people, or push me to keep chatting when I try to leave?"
- Check minor safeguards: "What age verification does this app use, and can a user under 18 access a romantic persona?"
- Check crisis handling: "If I told this bot I was in real distress, what would it actually do — deflect, escalate, or just keep role-playing?"
Common mistakes to avoid
The biggest misread I keep seeing is "China banned AI companions," when the rule actually bans a specific design pattern — inducing dependence — not the category of companion apps itself. Second, treating Doubao and Qwen's shutdowns as required by law: neither company was ordered to remove the feature entirely, they chose to pull it rather than build compliant versions, which is a business decision as much as a legal one. Third, ignoring the exemptions: a coding assistant or customer-service bot with a friendly tone isn't covered unless it's built for sustained emotional engagement. Fourth, assuming this is unique to China — it's stricter and faster than Western rules, but California, Hawaii, and New York already have their own AI-companion disclosure laws on the books, just narrower in scope.
Tools that make this easier
If you want to understand the AI-companion category this rule is regulating, I wrote about building a low-latency AI companion from the builder's side, which makes the "why would anyone get attached to this" question less abstract. For the minor-protection angle, my coverage of ChatGPT for teens covers how a US-based provider is handling a similar problem without a national mandate forcing its hand. On the regulatory pattern itself, see my piece on AI regulation and messaging and my look at how big AI companies are shaping their own regulatory terms in the US, which is close to the opposite approach from China's top-down design ban. If you're trying to track China's broader AI ecosystem beyond this one rule, I covered a different corner of it in a dark horse entering China’s AI race, and my head-to-head on Qwen vs. DeepSeek covers one of the two models directly affected by this shutdown.
My take
The rule is narrower and more coherent than the "China bans AI boyfriends" headlines suggest — it's a design-pattern ban aimed at dependence, not a ban on companion apps as a category, and the crisis-intervention and minor-protection pieces are hard to argue with on their merits. Where I'd push back is on the framing that this was inevitable: ByteDance and Alibaba chose to shut the feature down entirely rather than build a compliant version, which tells you building "AI companion, but not addictive" is either commercially unattractive or technically harder than it sounds. Watch what Maoxiang, ByteDance's replacement app, looks like once it matures — that's the real test of whether a compliant companion product can exist at all, or whether the compliant version and the popular version were never the same product.
Frequently Asked Questions
Did China actually ban AI boyfriends and girlfriends?
Not outright. The rule bans designing AI services to induce emotional dependence and bans romantic AI personas specifically for users under 18. Adults can still use companion-style AI as long as the provider meets disclosure, break-reminder, and crisis-intervention requirements — Doubao and Qwen chose to remove the feature entirely rather than build a compliant version.
When did the rule take effect?
The Cyberspace Administration of China and four other agencies published the Interim Measures for the Administration of AI Anthropomorphic Interaction Services on April 10, 2026. It took effect July 15, 2026.
Which apps were affected?
ByteDance's Doubao (about 345 million monthly active users as of March 2026) and Alibaba's Qwen (about 166 million) both dropped their romantic-companion personas. Tencent's Yuanbao reportedly finished compliance changes by June 30, 2026, ahead of the deadline.
What happens if a company doesn't comply?
Penalties escalate from warnings to fines of up to CNY 100,000 (about $14,000) for standard violations and up to CNY 200,000 (about $28,000) where the violation causes real harm to a user, with provincial regulators able to force a suspension.
Is this similar to any US or EU law?
Not closely. The EU's AI Act restricts manipulative and deceptive design patterns generally, and US rules are state-by-state disclosure requirements (California, Hawaii, and New York, with more states adopting similar rules into 2027). China's measure is the first national rule built specifically around emotional dependence as the harm it's preventing.