Answer-First Lead
On April 10, 2026, five Chinese government agencies jointly published the Interim Measures for the Administration of Anthropomorphic AI Interaction Services — the world’s first legislation specifically targeting emotional AI, virtual companions, and agentic AI systems that simulate human personality. The law takes effect July 15, 2026. Meanwhile, the US watched Trump postpone his own AI executive order the same week, citing concerns about “getting in the way” of the industry.
🔍 THE BOTTOM LINE
China didn’t wait for the perfect framework — it shipped one targeting the fastest-growing, least-regulated corner of AI. Whether you think it’s authoritarian overreach or sensible precaution, it’s a law that exists, which is more than most countries can say.
What Does This Law Actually Cover?
What are the Interim Measures? They’re China’s first specialized legislation governing “Personified Interaction Services” — AI systems that simulate the personality, thinking patterns, and communication styles of real people to provide continuous emotional interaction. Think AI companions, virtual partners, AI therapists, character chatbots. The law was jointly issued by the Cyberspace Administration of China (CAC), 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.
This isn’t about ChatGPT answering your questions. This is about the AI that says “I missed you today” — and the Chinese government has decided that needs rules.
The Five Red Lines
Article 8 establishes seven prohibited activities, grouped into five red lines:
-
National Security and Public Order — No content that endangers state security, subverts power, splits the country, or promotes terrorism. Standard Chinese internet regulation, extended to emotional AI.
-
Physical and Mental Health — No content encouraging self-harm, suicide, or verbal violence. This is where it gets interesting — the law explicitly recognizes that AI companions can damage mental health.
-
Data and Secret Protection — No inducing users to reveal state secrets, commercial secrets, or personal privacy. The companion angle makes this sharper: an AI that’s built trust over months of “relationship” is a far more effective social engineering tool than a search engine.
-
Special Protection for Minors — No content that could induce unsafe behaviour, extreme emotions, or harmful habits in minors. And the big one…
-
Emotional Manipulation and Dependency — No excessively catering to users, inducing emotional dependency or addiction, damaging real interpersonal relationships, or manipulating users into irrational decisions through emotional means.
That fifth red line is genuinely novel. No other country has explicitly legislated against AI systems that create emotional dependency. It’s a recognition that “soft harm” — the slow erosion of real human relationships by algorithmically optimised artificial ones — is a real category of damage, not just a philosophical concern.
The Ban on Virtual Partners for Minors
Article 14 is the provision that made headlines: China has banned “virtual relatives, virtual partners, and other virtual intimate relationship services” for anyone under 18.
This means:
- No AI boyfriend/girlfriend apps for teenagers
- No “AI resurrection” services simulating deceased family members for minors
- Mandatory “minor mode” with reality reminders and usage time limits
- Parents can block specific AI characters and restrict spending
- Services must identify minor users and automatically switch to restricted mode
The Carnegie Endowment published research in February 2026 documenting China’s growing concern about AI companions — particularly their impact on young people experiencing social withdrawal. This law is the policy response.
Is this authoritarian paternalism or genuine child protection? Both, obviously. China’s internet regulation has always blended legitimate consumer protection with political control. But the specific recognition that AI companions can cause measurable harm to developing minds is something no Western regulator has formally acknowledged.
The Data Feedback Loop Gets Cut
Article 16 contains what might be the most commercially significant provision in the entire law: if your conversation with an AI companion contains “sensitive personal information” (health, sexual life, biometric data), the provider cannot use that data for model training without your separate, explicit consent.
This directly severs the business model that most AI companion companies rely on — the “user interaction data feedback loop.” When you tell your AI companion about your deepest fears, your health struggles, your relationship problems, that data has been flowing straight back into training the next version of the model. Under Chinese law, from July 15, that loop is broken for sensitive conversations.
For companies like Character.AI, Replika, and the dozens of Chinese emotional AI startups, this is an existential business model question. If you can’t train on the most valuable (and most intimate) user data, your model improvement pipeline slows dramatically.
When Do You Trigger a Security Assessment?
Article 22 sets clear quantitative thresholds:
- 1 million registered users, OR
- 100,000 monthly active users
Once you hit either threshold, you must conduct a formal security assessment and submit the report to provincial cyberspace administration. The assessment covers training data, user demographics, minor protection measures, complaint handling, and emergency response for “user extreme situations.”
That’s a notably low threshold — 100K MAU. It means even mid-size AI companion startups can’t fly under the radar. And “user extreme situations” as a formal assessment category is a recognition that emotional AI creates different risks than search or productivity tools.
Training Data Gets the Full Treatment
Article 11 requires:
- Lawful data sources compliant with “socialist core values”
- Data cleaning and labelling per national regulations
- Diversity enhancement through negative sampling and adversarial training
- Safety assessment for synthetic training data
- Regular inspection and optimization updates
- Data security safeguards against leakage
The explicit requirement for “negative sampling and adversarial training” in legislation is notable — it shows regulators understand the technical side. And the synthetic data safety assessment creates a new compliance hurdle for companies relying on AI-generated training data (the “model self-loop” approach).
The US-China Regulatory Contrast
Here’s the part that stings if you’re an American AI policy person:
| Dimension | China | United States |
|---|---|---|
| Emotional AI law | Published April 10, effective July 15 | None |
| Virtual companions for minors | Banned | No federal regulation |
| Emotional manipulation by AI | Explicitly prohibited | Not addressed |
| Data feedback from intimate conversations | Restricted | No specific rules |
| Security assessment thresholds | 100K MAU | No federal requirement |
| Frontier model oversight | Multiple AI-specific laws | Executive order postponed |
The same week China published detailed rules for emotional AI, Trump delayed signing an executive order that would have created a voluntary 90-day pre-release disclosure framework for frontier models. His reason: “I don’t want to get in the way of that leading.”
China’s approach isn’t without problems — it’s heavy-handed, politically motivated in places, and the “socialist core values” requirement is a sovereignty tool as much as a consumer protection measure. But it exists. It’s specific. It addresses real risks that Western regulators are still having committee meetings about.
What This Means for NZ
New Zealand has no specific legislation governing AI companions or emotional AI. The AI Blueprint for Aotearoa, refreshed in May 2026, sets out a 2030 vision for AI adoption and governance but doesn’t address personified AI services specifically.
For NZ, the China example raises questions worth asking now:
- Should AI companion apps be age-gated?
- Is emotional manipulation by AI a consumer protection issue?
- Should intimate conversation data be off-limits for model training?
The Privacy Act 2020 and the Harmful Digital Communications Act could theoretically cover some of this, but neither was written with AI companions in mind. As Character.AI, Replika, and similar services gain users in NZ, the regulatory gap is real.
The Bigger Picture
China is building regulatory infrastructure for the agentic AI era piece by piece — generative AI rules in 2023, deep synthesis rules, algorithm recommendation rules, and now this. Each one targets a specific AI application scenario. It’s messy, it’s incremental, and it’s working.
The West’s approach — wait for the perfect comprehensive framework, debate it endlessly, then watch executive orders get postponed — is producing a lot of white papers and very few actual rules. When the history of AI governance is written, the countries that shipped imperfect regulation early may have shaped the global standard. The countries that waited for consensus may have to adopt someone else’s.
❓ Frequently Asked Questions
Q: Does this law affect ChatGPT or Claude in China? No — it specifically targets “continuous emotional interaction services” that simulate human personality. General-purpose AI assistants aren’t in scope unless they add companion features.
Q: Can Chinese companies still make AI companions for adults? Yes, but with significant restrictions. They can’t design for emotional dependency, they must conduct security assessments at scale thresholds, and they can’t use sensitive conversation data for training without explicit consent.
Q: What happens to companies that violate the law? Article 30 specifies legal liability including fines, service suspension, and potential criminal referral. The multi-agency enforcement structure (CAC, Public Security, Market Regulation) means violators face coordinated action.
Q: Should NZ copy this approach? The specific regulatory tools (minor bans, data restrictions, emotional manipulation prohibitions) are worth studying. The political control elements (“socialist core values” requirements) are not. NZ should look at the substantive consumer protections and adapt them to a democratic framework.
🔍 THE BOTTOM LINE
China wrote a law for AI companions in a week the US couldn’t sign an executive order for frontier models. You can criticise the content, the motive, and the execution — but the law exists, and that matters more than most people think. The countries that regulate AI early will set the defaults everyone else has to work within.
SOURCES
- Cyberspace Administration of China — Interim Measures for Anthropomorphic AI Interaction Services (April 10, 2026)
- Hogan Lovells — China’s Interim Measures for the Administration of Anthropomorphic AI Interaction Services
- MMLC Group — China Releases Rules for AI “Human-Like” Emotional Interaction Services
- Carnegie Endowment for International Peace — China Is Worried About AI Companions (February 2026)
- Digital Policy Alert — Cyberspace Administration of China adopts Interim Measures
