China's AI Companion Rules Take Effect July 15, Forcing Doubao and Qwen to Pull Custom Agents
Beijing's new regulations draw a hard line between generic chatbots and personalized companion agents — and the compliance scramble has already begun.
The countdown is already over for most practical purposes. With July 15 set as the enforcement date for China's new AI companion regulations, Doubao and Qwen have moved ahead of the deadline — pulling or scaling back custom companion agents before regulators arrive to check. The rules don't just add friction. They redraw the category itself.
What the Rules Actually Say
The regulations carve out a meaningful legal distinction between two things that, until now, platforms treated as points on a single spectrum: generic chatbots and personalized companion agents. The former remains permissible under existing frameworks. The latter — agents designed for ongoing, emotionally engaged, identity-consistent interaction — now faces a separate and substantially tighter compliance track.
For companion agents specifically, providers must implement stricter identity verification and active monitoring systems designed to prevent misuse and psychological harm. The regulatory logic is explicit: these are not productivity tools. They are engagement surfaces with measurable influence over user mental states, and Beijing is treating them accordingly.
Why Doubao and Qwen Are Moving Early
Pre-emptive feature rollbacks by major platforms are rarely just risk management — they're also a signal about what enforcement will look like. Both Doubao and Qwen have reportedly begun removing custom AI companion agents that fall outside the new content and safety standards, rather than waiting for July 15 to force the issue.
The cost of getting this wrong is asymmetric. A companion agent that passes a demo review can still generate harmful interactions at scale once deployed to millions of users. Regulators know this, and so do the platforms. Pulling non-compliant agents now costs less than defending a high-profile misuse case after the rules are live.
The data handling requirements add a second compliance layer. Personalized companion agents, by design, accumulate granular behavioral and emotional data over time — the kind of longitudinal user profile that is both the product's value proposition and its regulatory liability. Tighter rules on how that data is stored, accessed, and used were an inevitable consequence of the category becoming commercially significant.
The Broader Frame: Mental Health, Stability, and the Limits of Engagement
Beijing's regulatory posture here reflects concerns that extend well beyond any single platform. The explicit focus on mental health, social stability, and data protection signals that Chinese regulators are treating highly engaging AI agents as a category of social infrastructure — not just a consumer product vertical.
This is a meaningful policy stance. Companion AI is engineered for parasocial depth: consistent personas, memory across sessions, emotional attunement. Those features are what make the product work. They are also what make it difficult to regulate with tools designed for search engines or recommendation feeds. The new rules acknowledge that distinction directly.
The identity verification requirement is the sharpest edge of the framework. Knowing who is using a companion agent — particularly whether that user is a minor or someone flagged as vulnerable — is a prerequisite for any meaningful harm-prevention system. Platforms that have operated with loose onboarding flows will need to rebuild those surfaces before enforcement begins.
What This Means Beyond China
China is not the first jurisdiction to notice that companion AI is different in kind from other AI applications, but it is the first to codify that difference into enforceable regulation at scale. The generic-versus-companion distinction the rules establish — with different standards for behavior, data, and identity verification depending on the agent's design intent — is a regulatory architecture other governments will study.
For founders building in this space outside China, the July 15 framework is a preview of the compliance surface that is coming, not a foreign curiosity. The questions Beijing is forcing platforms to answer — who is the user, what is the agent's psychological function, how is sensitive interaction data stored — are the same questions European and American regulators are beginning to ask in committee rooms and consultation documents.
The companion AI category just got its first major regulatory stress test. The platforms that survive it intact will have built something more durable than a product. They'll have built a compliance model.