China's first-ever specialized law regulating emotional AI companions takes effect in just over a week, and the consequences are already hitting home. ByteDance's Doubao — China's most popular AI app with 330 million users — will shut down its personalized agent features on July 15 to comply with the new Interim Measures for the Administration of AI Personified Interactive Services. Alibaba's Qwen (Tongyi Qianwen) will follow suit the same day, and the two companies are handling user data very differently.
What the Law Actually Says
The regulation, finalized in April 2026 after a December 2025 draft period, is the world's first dedicated legal framework for "anthropomorphic AI" — systems that simulate human personality, thinking, and communication styles while engaging in continuous emotional interaction. It takes effect on July 15, 2026, and applies to any domestic public service offering such interaction.
Key provisions targeting the companion space include:
- A complete ban on offering "virtual partner" or "virtual relative" services to minors under 18
- Requirements to prevent emotional dependency, addiction, or the replacement of human social interaction
- Mandatory AI disclosures with reminders every two hours of continuous use
- A prohibition on using conversation data (including chat logs about health, sexual life, or biometric details) for model training without explicit opt-in consent
- Human operator intervention required when users show signs of self-harm
Violations can result in service suspension and fines up to 200,000 RMB (~$28,000). Providers serving over 1 million registered users must complete security assessments and file with provincial cyberspace authorities before launching new anthropomorphic functions.
Doubao: Agent Features Gone, Core Chatbot Remains
ByteDance announced on July 4 that Doubao will disable its AI agent functionality — the part of the app that let users create personalized, autonomous agents capable of booking flights, managing workflows, and executing multi-step tasks. The standard chatbot remains operational.
The data situation for Doubao users: They have until the July 15 shutdown to export their personalized agent data. A read-only grace period for agent configurations and chat histories runs until October 15, 2026. After that, all Doubao agent data will be permanently deleted and unrecoverable.
ByteDance is directing users to its alternative "Cat Box" (Mao Xiang) app for similar agent functionality.
Qwen: Less Generous with User Data
Alibaba's Qwen is taking a harder line. Its agent features also shut down on July 15, but the company has announced no grace period or data migration path. Personalized agent data will be permanently deleted on shutdown day with no read-only window.
The discrepancy highlights a broader tension: persistent-memory AI agent architectures — where agents remember user history, preferences, and relationships over time — are structurally at odds with China's new anti-addiction and data protection mandates. The regulation treats emotional manipulation by anthropomorphic AI as a legal harm for the first time globally.
Global Precedent for Companion Regulation
China's move comes as companion AI faces increasing regulatory scrutiny worldwide. California's SB 243, which mandates companion chatbot safety and disclosure rules, has been in effect since January 2026. The EU AI Act's Article 50 transparency rules for chatbots become fully enforceable on August 2, 2026.
But China's approach is the most aggressive yet. Where SB 243 focuses on disclosure and crisis prevention, and the EU AI Act on transparency labels, China's law goes directly after product design: it bans the emotional architectures that make companion AI sticky. The regulation explicitly prohibits designing services to "replace social interaction," "control user psychology," or "induce addiction."
For global companion app developers, this is a signal. If the world's largest AI market is banning virtual partners for minors and forcing companies to dismantle memory-based agent systems, similar logic could appear in other jurisdictions. The law's definition of harm — emotional dependency as a legally recognized injury from AI — is a framing that companion companies should expect to see elsewhere.