As of today, running an AI that acts like your friend is a regulated activity in China. The Interim Measures for the Administration of AI Anthropomorphic Interactive Services — issued April 10 by the Cyberspace Administration of China and four partner agencies — took effect July 15. The country’s two biggest consumer AI apps didn’t wait to find out how enforcement feels.
ByteDance’s Doubao took its custom agent feature offline today; users can view old configurations and conversations in read-only mode until October 15, after which the data becomes unrecoverable. Alibaba’s Qwen disabled user-created humanlike agents on July 10 and pulled its broader agent functions today. Millions of AI companions — digital boyfriends, virtual family members, custom personas — went dark in a single week.
Read the scope clause and the design is obvious. The rules cover services that “simulate human personality traits, thinking patterns and communication styles” to provide sustained emotional interaction. Productivity assistants, Q&A bots, and customer-service systems are explicitly exempt. Beijing didn’t regulate AI. It regulated the relationship — and left the same models free to power the open-source workhorses China is exporting to the world.
Our take: Ignore the fines — 100,000 yuan is about $14,700, pocket change for ByteDance. The real penalty is product death: regulators can suspend registrations and block features, which is why both giants amputated the category before day one. The bigger signal is the split itself. China just became the first jurisdiction to draw a legal line between AI-as-tool and AI-as-relationship, regulating only the second. Companion AI is the stickiest consumer category in the business — and every Western lab building “memory” and “personality” into assistants is drifting toward the regulated side of that line. When Washington and Brussels finally move on companion bots, this framework is the template they’ll be handed. Compliance is also a moat: minor modes, distress detection, and escalation systems are infrastructure only big platforms can afford.
The fine print that bites
The measures ban offering virtual companions or virtual family members to minors outright, require guardian consent for users under 14, and mandate dedicated minor modes with usage-time limits and prompts to return to the real world. Providers must also detect users in acute distress — signs of self-harm, suicidal behavior, or serious financial loss — and escalate to guardians or emergency contacts. That’s not a content rule; it’s a duty-of-care standard, the first imposed on AI anywhere.
Enforcement didn’t wait for today either. Shanghai’s internet regulator said on June 26 it had already removed more than 14,000 non-compliant AI agents, citing impersonation of officials, vulgar role-play, and unauthorized data collection. The sweep came while 193 countries in Geneva were still debating whether to write any binding AI rules at all. China skipped the communiqué stage.
What to watch
- October 15: Doubao’s data-deletion deadline. Watch for a compliant “licensed companion” relaunch before then — the demand didn’t disappear, the product did.
- The gray zone: assistants with persistent memory and personality sit between “tool” and “relationship.” Where regulators file them decides how much of China’s booming consumer AI layer needs a license.
- Western echoes: US state legislatures and the EU have companion-bot bills in draft. The tool/relationship distinction is now sitting there, pre-written.