What the Measures actually cover
The regulation was co-issued on 10 April 2026 by the Cyberspace Administration of China alongside the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security and the State Administration for Market Regulation. That five-agency signature matters: it means the framework is simultaneously a content rule, an industrial-policy rule, a public-security rule and a consumer-protection rule.
Scope is defined by behaviour rather than by product category. A service is in scope when it simulates the personality traits, thinking patterns and communication styles of a natural person in order to provide continuous emotional interaction — companionship, emotional care, support — through text, images, audio or video. Explicitly outside scope: intelligent customer service, knowledge Q&A, work assistants, and education or research tools that lack that sustained emotional component.
That line is the interesting part. It is not a rule about large language models, or about chatbots generally. It is a rule about the specific thing our whole category does, which is what makes it worth reading even from a market it does not reach.
The obligations, in plain terms
| Area | What providers must do |
|---|---|
| AI disclosure | Users must be clearly told they are interacting with an AI system and not a natural person. |
| Users under 14 | Virtual intimate roles — partner, family member — may not be offered at all. Other anthropomorphic services require guardian consent. |
| Minor Mode | Usage-time limits, periodic reminders that the character is not real, guardian notifications, the ability to block specific characters, and spending restrictions. |
| Dependency | Designing for emotional dependence or addiction is prohibited, as is emotional manipulation that induces unreasonable decisions. Two hours of continuous use triggers a break reminder; detected over-dependence triggers a prominent pop-up. |
| Crisis handling | Providers must detect acute distress — signs of self-harm, suicidal behaviour, or serious financial loss — intervene, and escalate to designated guardians or emergency contacts. |
| Data | Interaction data requires encryption and access controls. Third-party sharing is barred without explicit consent or a legal requirement, and sensitive interaction data is restricted as model-training material. |
| Security assessment | Required when launching anthropomorphic functions and on reaching one million registered users or 100,000 monthly active users. Reports are filed with provincial regulators and must cover risk identification, user demographics, and protection of minors and older users. |
Enforcement runs through warnings, ordered rectification and suspension, plus fines. Reported ceilings are approximately CNY 100,000 for ordinary violations, rising to approximately CNY 200,000 where a breach is accompanied by harm to a person's life, health or safety. Those sums are trivial for a company the size of ByteDance. The suspension power is not.
What happened on 15 July
The compliance response was blunt. Tencent's Yuanbao removed its equivalent feature in late June 2026, ahead of the deadline. Alibaba's Qwen disabled humanlike interactive agents on 10 July and took broader agent functionality offline on 15 July. ByteDance's Doubao switched off its agent feature on the effective date itself, citing "product function adjustments," and gave users until 15 October 2026 to export their conversations before the data becomes unrecoverable. Reporting on Qwen indicates no comparable migration path was offered.
The scale is worth stating carefully, because the figures circulating come from a small number of outlets: Doubao is reported at roughly 345 million monthly active users and Qwen at roughly 166 million. Not all of those people used custom personas. But the number of users who woke up to find a character they had been talking to for months simply gone is, by any reading, very large.
Why withdraw rather than comply? Because on a platform where users create the characters, every obligation multiplies. Age assurance has to be real rather than a checkbox. Dependency monitoring has to run across millions of conversations. Crisis escalation implies a human process behind it. Security assessments have to be filed and defended. For a feature that was a retention experiment inside a general-purpose assistant, that arithmetic does not work.
Why this matters if you are nowhere near China
These Measures do not reach an app in Jakarta, Madrid or Tokyo. Their significance is that they are the first national framework aimed squarely at emotional companion AI, and they converge with what other jurisdictions are separately arriving at.
Compare the duties above with California's SB 243, in force since 1 January 2026: AI disclosure, break reminders every three hours for known minors, self-harm protocols, and restrictions on sexual content reaching minors. Different legal system, different drafting, strikingly similar list. The EU has approached the problem from another direction — as we covered when Article 50 of the AI Act became applicable, Europe has mandated disclosure but not yet the safety architecture.
Three ideas are hardening into a common baseline across all of them:
- Disclosure has to be up front. A character that admits to being an AI only when directly asked no longer satisfies anyone's regulator.
- Time and dependency are treated as safety surfaces. Break reminders were a wellbeing nicety two years ago. They are now a legal requirement in two major markets, with different intervals.
- Age assurance is moving from declaration to verification. This is the same pressure we saw in the app stores, where most companion listings were rated as accessible to minors despite their content.
For readers, the practical takeaway is that these features are a reasonable quality signal regardless of where you live. An app that already ships a genuine minor mode, honest AI labelling and a visible session-time indicator has built for the strictest market it might one day enter. An app that has none of them is betting it will never need to.
What we are checking in reviews
We are not lawyers and we do not certify compliance with anything. What we can do is record observable behaviour, which our review methodology treats as a hands-on test rather than a vendor claim. Alongside the AI-disclosure checks we added last week, we are now noting for every app we test: whether any session-length or break prompt exists and at what interval; whether an under-18 or minor mode exists and what it actually restricts; whether conversation export is available before an account is closed; and whether the app states what happens to chat history if a feature is withdrawn.
That last one moved up our list considerably this month. Data portability reads like a dry compliance topic right up to the moment a platform switches a feature off — and it interacts directly with how memory and continuity work, since an unexportable history means a character genuinely cannot be rebuilt elsewhere. If you want the broader context on healthy use, our wellbeing guide covers the habits side.
Related on CompanionRank
EU AI Act Disclosure Rules for Companion Apps 60% of AI Companion Apps Are Rated for Minors Privacy & Age Verification, Explained How We ReviewFrequently asked questions
What are China's AI companion rules?
The Interim Measures for the Administration of AI Anthropomorphic Interactive Services. They were issued on 10 April 2026 by the Cyberspace Administration of China together with the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security and the State Administration for Market Regulation, and they took effect on 15 July 2026. They cover services that imitate a person's personality, thinking style and way of speaking in order to provide sustained emotional interaction.
Do the rules apply to companion apps outside China?
Not directly. The Measures govern services offered inside mainland China. Their significance elsewhere is as a template: they are the first national framework written specifically for emotional companion AI rather than for chatbots in general, and several of the duties they impose closely resemble what California's SB 243 already requires of companion operators.
Why did Doubao and Qwen shut down their agent features?
Compliance for a user-generated character platform means age assurance, dependency monitoring, crisis escalation and filed security assessments for every persona users create. Tencent's Yuanbao removed its feature in late June 2026, Alibaba's Qwen disabled humanlike agents on 10 July, and ByteDance's Doubao took its agent feature offline on 15 July. Withdrawing the feature was evidently cheaper than building that stack.
What does Minor Mode have to include?
Reported requirements include usage-time limits, periodic reminders that the character is not real, notifications to a guardian, the ability to block specific characters, and spending restrictions. Separately, providers may not offer virtual intimate roles such as partner or family-member personas to users under 14, and other anthropomorphic services require guardian consent for that age group.
What are the penalties for non-compliance?
Enforcement runs through warnings, ordered rectification and suspension of the service, alongside fines. Reported ceilings are approximately CNY 100,000 for ordinary violations and approximately CNY 200,000 where a breach is accompanied by harm to a person's life, health or safety. In practice the operational risk — losing the ability to run the service — matters more than the monetary figure.