China Bans Humanized AI Companions, Forces ByteDance and Alibaba Shutdowns

Published by James Harris on July 6, 2026July 6, 2026

China’s two largest AI platforms are pulling their companion and custom agent features within days of each other, not because users stopped caring but because a new regulatory framework gives them no choice.

ByteDance’s Doubao will disable its custom AI persona function on July 15, with all user data either deleted or archived by October 15. Alibaba’s Qwen is moving even faster, removing human-like bots on July 10 and shutting agent functions on July 15. Tencent’s Yuanbao had already moved ahead of both. The trigger is China’s “Interim Measures for the Administration of Humanized Interactive Services Based on Artificial Intelligence,” proposed in April and effective July 15, which targets AI that simulates human personality traits, cognitive style, and prolonged emotional relationships. The law was explicitly motivated by concerns over emotional dependency, damage to minors’ mental health, data security, and what legislators described as ethical compromises. Platforms are now prohibited from using personal chat history to train future models, a provision that carries significant implications for how these companies have historically monetized their most engaged users.

The carve-outs matter as much as the restrictions: workplace assistants, customer service bots, educational tools, and research platforms are exempt, provided they do not facilitate prolonged emotional engagement.

ByteDance’s clarification that its Cat Box App will retain agent-creation features suggests the industry is already learning to route around the new rules rather than comply wholesale. The companion AI market was always a different product category from productivity tools, but regulators drawing that line formally is new, and the distinction will now shape product roadmaps across the sector. This also lands in a context where access to American AI products in the region has been determined largely by individual companies rather than a unified framework, making China’s top-down legislative move a structural contrast worth watching. Globally, OpenAI and Character.AI are facing high-stakes US lawsuits over similar emotional dependency concerns, and the Federal Trade Commission has opened inquiries into xAI, Snap, and Character Technologies over risks to children. China has simply moved from litigation pressure to legislation first.

The deeper question is whether the exemption architecture holds. Regulators wrote the carve-outs broadly enough that a determined platform could rebuild most companion functionality under an “educational” or “productivity” label. If that happens, a second round of enforcement is probably more restrictive than the first.