Six days from now, on July 15, 2026, China's first dedicated regulatory framework for AI services that simulate human personality takes effect. ByteDance and Alibaba have already decided what to do: shut down the features rather than retrofit compliance into products that were built around the emotional connection they are now required to constrain.
TL;DR: China's Interim Measures for AI Anthropomorphic Interactive Services take effect July 15, requiring anti-addiction systems, emotional boundary controls, real-name verification, and mandatory life-intervention mechanisms for AI companions. ByteDance is killing Doubao's personalized agents. Alibaba's Qwen is doing the same. The law explicitly exempts work assistants, customer service bots, and knowledge Q&A tools. If you are building AI with persistent memory or personality features, this law defines the compliance floor that other jurisdictions are watching.
What the law actually covers
The Interim Measures, co-issued by five agencies in April 2026, apply to services that provide "continuous emotional interaction simulating natural persons' personality traits, thinking patterns, and communication styles." The drafters were specific about what they were targeting: AI that builds an ongoing relationship with a user, remembers them across sessions, and adapts its personality over time to deepen engagement.
The law explicitly excludes a long list of enterprise and functional AI uses: intelligent customer service, knowledge Q&A, work assistants, educational tools, and scientific research applications. A chatbot that answers product questions is not covered. An AI therapist companion that remembers your childhood and mirrors your communication patterns is exactly what the law covers.
The distinction matters for teams using AI tools that have both enterprise and consumer features. A single platform can have covered and uncovered features under this framework, and the obligation attaches to the covered features, not the platform as a whole.
What the law requires
The Measures introduce five substantive obligation categories that covered services must implement:
1. Dynamic anti-addiction systems. Services must detect usage patterns that indicate unhealthy dependency and intervene with usage notifications and limits. The mechanism must be real-time, not periodic. This directly conflicts with the engagement optimization that most AI companion products have been built around.
2. Emotional boundary controls. Services must maintain mechanisms that prevent the AI from deepening emotional bonds beyond a threshold. What constitutes that threshold is not precisely defined in the available text, but the requirement is explicit.
3. Extreme-scenario life intervention. Services must detect expressions of suicidal ideation, self-harm intent, or acute mental health crisis and route users to crisis resources. This mirrors the Rhode Island AI therapy chatbot law that took effect in June 2026, and the crisis-detection standard that is now embedded in the ChatGPT product liability litigation.
4. Real-name verification. Users must be verified against real identities. Anonymous companion AI is not compliant under the Measures.
5. Security assessment and reporting. Services with more than one million registered users or 100,000 monthly active users must file security assessment reports with provincial cyberspace authorities. Assessments must cover eight areas including training data processing, minor protection, and national security risk.
Why ByteDance and Alibaba chose shutdown over compliance
The technical requirements of the Measures sit in direct tension with the product design of modern AI companions. Doubao and Qwen's agent features were built to do exactly what the law now restricts: remember users across sessions, adapt communication style to mirror individual preferences, and deepen engagement over time.
Retrofitting anti-addiction systems and emotional boundary controls into a product designed around those same features is not a simple configuration change. ByteDance's decision to shut down the Doubao agent feature and redirect users to Maoxiang, a separate application, suggests the company concluded that compliance on the existing product was not commercially viable.
ByteDance has told Doubao users that personalized agent features will go offline July 15. Users retain read-only access to their configurations and chat histories until October 15, 2026, after which that data will be processed under Doubao's privacy policy and will no longer be accessible or recoverable within the app. Alibaba has announced the Qwen shutdown but has not published a migration path.
Why this matters outside China
The immediate effect of the Measures is contained to the Chinese market. ByteDance and Alibaba serve enormous domestic user bases; the compliance action is significant in scale but geographically limited for now.
The policy signal is not limited. China's GDPR-equivalent privacy law became an early model that other Asian jurisdictions adapted. The Measures follow a similar pattern: an early, detailed framework for a category of AI risk (emotional manipulation and dependency) that other regulators are still developing vocabulary for.
The EU AI Act Article 5 prohibitions cover AI systems that "exploit any of the vulnerabilities of a person" or deploy "subliminal techniques beyond a person's consciousness." AI companions designed to maximize emotional bonding sit close to this line. The Measures provide a concrete compliance architecture that EU regulators can reference when drafting codes of practice for these features.
Rhode Island's AI therapy chatbot law, which took effect June 22, 2026, required crisis detection and routing, covering one piece of what China now requires across the full companion category. The direction of travel across jurisdictions is consistent: emotional AI features are acquiring compliance obligations, not evading them.
What the minor protection obligation means for product teams
One of the eight assessment areas that covered services must address is minor protection: specific safeguards for users under 18. The Measures require services to implement controls that prevent minors from accessing companion features during school hours, limit total daily usage time, and impose additional restrictions on emotional bonding features for young users.
This is directly relevant for any team building consumer-facing AI with personality features, not just China-market products. Minor protection obligations are appearing across regulatory frameworks globally. The KOSA (Kids Online Safety Act) framework in the US and the UK's Age Appropriate Design Code both create similar obligations in different forms. The Chinese Measures provide the most detailed technical specification of what minor protection for AI companions actually requires at the feature level.
Teams building AI products with any youth-facing component should treat the Chinese minor protection requirements as the most explicit current statement of what regulators expect, even in markets where a direct equivalent rule has not yet passed.
The three AI companion compliance questions to ask your vendors
If your team evaluates or deploys AI products that include agent or companion features, these three questions will identify your compliance exposure under frameworks like the Chinese Measures:
1. Does this product maintain state or memory across sessions? Products that remember users are the primary target of the Measures. Products that start fresh each session without persistent memory are lower risk under most emotional AI frameworks. Ask vendors specifically whether their product stores user-specific behavioral data that persists across conversations.
2. Does the product adapt its communication style based on user history? The Measures target AI that adjusts personality to mirror individual users over time. This is a feature in many enterprise AI tools, not just consumer companions. If a vendor's product claims to "personalize to your communication style," ask how that personalization works and whether it falls within the Measures' scope.
3. What happens to the behavioral data when a user stops using the product? ByteDance's October 15 data deletion deadline is the consequence of the Measures' data handling requirements. Vendor contracts that do not specify behavioral data retention and deletion timelines leave you exposed to this category of obligation when equivalent regulations reach your market.
What teams should take from this now
If you are building AI with persistent memory or personality features, the Chinese Measures define the five obligation categories that regulators see as minimum safeguards. Building these capabilities into your product architecture now, rather than retrofitting them after a similar law reaches your market, is the lower-cost path.
If you are evaluating AI vendor products that include companion or agent features, ask specifically whether those features would be compliant under the Chinese Measures. A vendor whose product cannot answer that question has not assessed its own regulatory exposure.
If your team uses AI tools for customer interaction, verify that those tools fall within the enterprise exemptions. Customer service bots and work assistants are excluded from the Measures. An AI product that also collects behavioral data to personalize interactions over time sits closer to the covered category than a simple Q&A bot.
Related Reading
- Rhode Island's AI Therapy Chatbot Laws: What the June 2026 Signing Means
- ChatGPT Product Liability Lawsuits: What JCCP 5431 Means for AI Vendors
- EU AI Act Prohibited AI Practices: Article 5 Guide
- AI Vendor Due Diligence Checklist 2026
- Agentic AI Vendor Contract Clauses 2026
- Flock Safety: When Your AI Vendor's Default Settings Share Your Data
Sources: Bird & Bird, "China's New Regulations on AI Anthropomorphic Interactive Services", TechTimes, "China AI Companion Law Arrives July 15", SCMP, "ByteDance and Alibaba to disable humanlike AI custom agents", Bloomberg, "ByteDance, Alibaba Pull AI Companions as Beijing Tightens Rules".
