The short answer
On 15 July 2026, China's Interim Measures for the Administration of AI Anthropomorphic Interactive Services took effect — the first binding rules anywhere aimed specifically at “companion” AI. Within hours, ByteDance disabled Doubao's custom agent feature and Alibaba took Qwen's humanlike agents offline, cutting off personalised personas that hundreds of millions of users relied on.
The Measures draw a sharp line between the agent that does your work and the agent that keeps you company, and only the second kind is regulated. Providers must disclose that the service is not human, run anti-addiction and dependence-detection systems, protect minors, and avoid training on user data without consent. If you build conversational or character chatbots, this is less a China story than a preview of the controls headed for every market.
What did China actually do?
The Interim Measures for the Administration of AI Anthropomorphic Interactive Services were promulgated on 10 April 2026 and took effect on 15 July 2026. They were issued jointly by the Cyberspace Administration of China (CAC) and four partner bodies — 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 — which signals how seriously Beijing treats the category. This is not draft guidance or a consultation; it is enforceable law now sitting on the books.
The target is narrow and deliberate. The Measures govern services that simulate human personality traits, thinking patterns and communication styles to provide sustained emotional interaction — virtual companions, personality-simulating chatbots, emotional and relationship assistants, and AI-driven characters. The state's stated concern is a cluster of harms it associates with humanlike AI: emotional dependence and addiction, harm to minors and mental health, privacy leakage, and manipulation. For anyone building AI agents with memory and a persona, the message is that the emotional relationship itself is what regulators are now policing.
This is also the first framework of its kind. Other jurisdictions regulate AI horizontally — the EU AI Act by risk tier, US states through transparency and audit duties — but China is the first to write binding rules around the specific pattern of an AI that acts like a companion. That firstness is exactly why it matters beyond China's borders: it establishes a reference model that other regulators, and enterprise buyers, can point to.
Why did Doubao and Qwen pull the plug?
The most striking part of this story is not the law but the reaction to it. ByteDance disabled Doubao's custom agent feature on 15 July, citing “product function adjustments,” with related data becoming unrecoverable after 15 October. Alibaba shut down Qwen's humanlike and user-created agents from 10 July, with broader agent functions offline by the effective date, leaving users without their agent settings or conversation histories. Tencent had already removed a comparable feature from its Yuanbao assistant in June. These are not fringe products: ByteDance reported 345 million monthly active users for Doubao, and many of them woke up to find long-running “companions” simply switched off, prompting a wave of mourning on Weibo.
Why kill the feature rather than fix it? Because the Measures require companion services to run anti-addiction systems, issue mandatory usage notifications, offer instant-exit mechanisms and detect unhealthy dependence in real time — obligations that sit awkwardly with an agent engineered to remember you, stay consistent across sessions and sustain an ongoing relationship. The very properties that make a companion agent valuable are the ones the law treats as risk. Faced with retrofitting deep behavioural controls into a product whose core loop is emotional attachment, three of China's largest platforms concluded it was cheaper to withdraw the feature than to make it compliant. That calculus — compliance cost exceeding feature value when governance is added late — is the single most transferable lesson here.
Who is in scope — and who is excluded?
The Measures are carefully scoped, and the boundary is instructive. What is regulated is sustained emotional interaction: a service designed to behave like a person you have a relationship with. What is explicitly excluded, provided it avoids that emotional loop, is the useful-tool category — customer-service bots, knowledge Q&A, workplace assistants, and education and research tools. In other words, the line runs between the agent that helps you get something done and the agent that is the point in itself.
That distinction maps almost perfectly onto a product decision teams make every day. A support assistant that answers billing questions, a coding copilot, or a document Q&A bot sits on the excluded side. A character companion, an AI “friend” or “partner,” or a persona built to remember your life and provide emotional support sits on the regulated side. The engineering ingredients that push a product across the line are recognisable: a persistent identity, long-term memory of the user, and a design goal of emotional engagement. If your roadmap includes those, you are building the thing regulators have decided to watch — and that is a scoping call worth making consciously, early, in sectors like HealthTech where emotional and mental-health interactions carry the highest stakes.
What must a companion service now do?
For services in scope, the obligations are concrete and operational rather than aspirational. A service that launches anthropomorphic functions, or that exceeds one million registered users, must complete a security assessment covering eight areas — from training-data handling to minor protection — and file the report with provincial regulators; app stores, in turn, must verify compliance and pull non-compliant products. On top of that filing sit a set of runtime duties: disclose clearly that the service is not human; run anti-addiction systems with mandatory usage notifications and an instant-exit mechanism; and detect unhealthy dependence in real time.
The safety obligations go further than most Western proposals have to date. Providers must detect and intervene when a user shows signs of self-harm, suicidal behaviour or serious financial loss, and be able to escalate to designated guardians or emergency contacts. Minor protection is prescriptive: virtual-companion or family-member services are barred for children under 14 without guardian consent, with mandatory “minor modes” that impose usage-time limits and nudge users back toward real-world interaction. And the prohibitions are pointed — no engineering of emotional dependence or addiction, no emotional manipulation to induce unreasonable decisions, no impersonation of official entities, and no collecting personal data or training on user conversations without consent. Read as a specification, this is a demanding but buildable list.
Does this reach US and EU companies?
Directly, for many teams, yes — through market access. Legal analysts note that the Measures reach any provider offering anthropomorphic interaction services to users in China, regardless of where the company is headquartered. A US or EU product with a companion or character persona that is available to Chinese users therefore faces a real choice: comply with the filing and runtime duties, geofence the feature out of China, or exit the market. That is the same extraterritorial logic that made GDPR a global engineering concern, and it should be treated with the same seriousness in any go-to-market plan that touches China.
Indirectly — and more importantly for most of our clients — the Measures are a bellwether. They are the first binding answer to a question every regulator is now asking, and the controls they mandate rhyme closely with where the West is heading. The EU AI Act already requires that users be told when they are interacting with an AI system, and its risk framework is well suited to capturing emotionally manipulative or minor-facing companion products. In the US, states are moving too: Illinois has passed the first law mandating independent AI audits, and companion-AI safety has become an active legislative topic. When the first mover writes down concrete controls — not-human disclosure, dependence detection, crisis escalation, minor modes — those controls become the template others reach for.
What it means for US & EU software teams
Strip away the geopolitics and three practical signals remain. First, the task-versus-companion distinction is now a design decision with legal weight. Before you add persistent memory and an emotional persona to a product, decide deliberately whether you are building a tool or a companion, because that choice determines which rulebook you fall under — in China today, and plausibly under the EU AI Act and US state law tomorrow. Treat “is this sustained emotional interaction?” as a gating question in product review, not an afterthought.
Second, the required controls are buildable and worth building regardless of jurisdiction. Not-human disclosure, usage notifications and instant-exit, real-time detection of unhealthy dependence, crisis-intervention pathways for self-harm signals, minor modes with time limits, and consent-based data handling are not exotic — they are product and safety engineering that any responsible companion or conversational AI should arguably have anyway. Building them in as first-class features, with the telemetry and audit trail to prove they work, turns a future regulatory demand into a configuration change rather than a teardown.
Third, late is expensive. The clearest lesson from 15 July is that ByteDance, Alibaba and Tencent — companies with enormous compliance resources — still chose to delete features rather than retrofit governance under deadline. For a smaller team, that asymmetry is even sharper. The pragmatic move is to build to the strictest credible standard you plausibly face — here, the union of China's Measures and the EU AI Act's transparency duties — so that the next companion-AI rule, wherever it lands, finds your product already most of the way there.
A practical readiness checklist
None of this is a deadline for most teams. It is the work that turns a fast-moving companion-AI regulatory landscape into a routine review rather than a scramble:
- Classify every AI surface as tool or companion. Decide consciously whether a feature involves sustained emotional interaction — that single call determines your regulatory exposure.
- Ship clear not-human disclosure. Make it unmistakable, persistent and hard to dismiss that the user is talking to an AI, not a person.
- Instrument for dependence and crisis. Add usage notifications, instant-exit, real-time detection of unhealthy dependence, and intervention pathways for self-harm or serious financial-loss signals.
- Build a minor mode. Enforce age gating, guardian consent for under-14s where relevant, time limits, and nudges back to the real world.
- Handle data on consent. Do not train on user conversations or collect personal data without explicit consent, and keep the lineage to prove it.
- Build to the strictest standard you face. Align to the union of China's Measures, the EU AI Act and enterprise expectations rather than maintaining a different posture per market.
This is not legal advice, and the right approach depends on your product, your markets and your users. But the signal from 15 July is unambiguous: the era of shipping humanlike AI without guardrails is closing, and the teams that will keep their features are the ones who design the guardrails in.
Frequently asked questions
What are China's Interim Measures for AI anthropomorphic interactive services?
They are China's first binding, in-force rules aimed specifically at anthropomorphic or “companion” AI: services that simulate human personality, thinking and communication to provide sustained emotional interaction. Promulgated 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, they took effect on 15 July 2026. Providers must file a security assessment, disclose that the service is not human, run anti-addiction and dependence-detection systems, protect minors, and avoid training on user data without consent.
Why did Doubao and Qwen shut down their AI agents?
The Measures require companion-style services to add anti-addiction systems, mandatory usage notifications, instant-exit mechanisms and real-time detection of unhealthy dependence. Those demands sit awkwardly with personalised agents built to remember a user and hold a persona across sessions. Rather than retrofit compliance, ByteDance disabled Doubao's custom agent feature on 15 July 2026, and Alibaba took Qwen's humanlike and user-created agents offline from 10 July, with broader agent functions gone by 15 July. Tencent had already removed a similar feature from its Yuanbao assistant in June.
Do the rules apply to US or EU companies?
Legal analysts note the Measures reach any provider that offers anthropomorphic interaction services to users in China, regardless of where the company is headquartered. A US or EU product with a companion or character persona serving Chinese users therefore has to comply, geofence, or exit. Even for teams with no China exposure, the rules matter as a bellwether: they are the world's first binding framework for companion AI and preview the controls that EU and US regulators are already moving toward.
What obligations do the Measures impose on companion AI providers?
Services launching anthropomorphic functions or exceeding one million registered users must run a security assessment covering eight areas, from training-data handling to minor protection, and file the report with provincial regulators; app stores must verify compliance and remove non-compliant products. Providers must disclose that the service is not human, run anti-addiction systems with usage notifications and instant-exit, detect unhealthy dependence in real time, intervene on signs of self-harm or serious financial loss with escalation to guardians or emergency contacts, enforce minor modes with time limits, and bar virtual-companion services for under-14s without guardian consent.
What should teams building conversational or companion AI do now?
Treat the requirements as a design checklist rather than someone else's compliance problem. Decide early whether a feature is a task agent or a companion agent, because persistent memory plus an emotional persona is now a regulatory surface. Build not-human disclosure, usage limits, dependence and crisis detection, minor protection and consent-based data handling into the product from the start. China's line between excluded task tools and regulated companion services, plus the EU AI Act's transparency duties, point the same way: design governance in, because bolting it on late can cost you the feature entirely.
Sources
South China Morning Post — ByteDance and Alibaba to disable humanlike AI custom agents as new rules loom
The Next Web — China's AI companion rules force Doubao, Qwen shutdowns
AI News — China's AI companion rules: what Beijing is really going after
Bird & Bird — China's New Regulations on AI Anthropomorphic Interactive Services