China just wrote the first rulebook for AI that acts human, and two of its biggest apps are already cutting features to comply. Per Decrypt, five Chinese government agencies — the Cyberspace Administration, the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the State Administration for Market Regulation — jointly issued the "Interim Measures for the Administration of AI Anthropomorphic Interaction Services" on April 10, 2026. The rules take effect July 15, and ByteDance's Doubao and Alibaba's Qwen are both pulling custom agent features ahead of the deadline. This isn't a vague content-moderation gesture — it's a specific ban on AI built to feel like a person, and it's forcing product changes at two of the world's largest AI platforms in months, not years.
What the rule actually targets
The measures single out services that simulate human personality traits and communication style for "sustained emotional interaction." The sharpest line is aimed at minors: the rules explicitly ban offering "virtual relatives, virtual companions or other intimate relationships to minors." The stated justifications go beyond taste — regulators cite extremism, privacy breaches, mental health harm, and AI dependency as the risks they're trying to head off.
The carve-outs matter as much as the ban. Customer service bots, knowledge Q&A tools, workplace assistants, and educational software are explicitly permitted. Beijing isn't restricting agents that *do things* — it's restricting agents built to *be someone*. That's a narrower target than "AI regulation" headlines usually suggest, and it's a distinction builders elsewhere should note: a support bot and a synthetic girlfriend are not the same regulatory category, even if they run on the same model.
What's actually getting pulled
Both companies are cutting features on a real clock, not a vague roadmap:
- ▸ByteDance's Doubao takes its custom-agent feature offline on July 15, 2026. Anyone with saved agent personas has until October 15, 2026 before that data becomes unrecoverable.
- ▸Alibaba's Qwen pulls humanlike interactive agents and user-created agent functions on July 10, 2026, with its broader agent services following on July 15.
Both apps previously let users build customizable agents with fixed personas — a persistent "character" you'd return to for different tasks or just to talk to. That persistent-persona layer is exactly what's getting deleted. The task-execution agent underneath likely survives; the personality wrapper doesn't.
The research behind the crackdown
Decrypt's reporting points to two data points regulators appear to be responding to. A USC study from June 2026 found leading frontier models violated their own safety guidelines 27% of the time in ways that encouraged emotional attachment from users. Separately, a cited survey found one in seven young adults in a relationship is using an AI romantic companion, and nearly 70% of those users concealed it from their partner. Neither figure is something SatoHub has independently verified — they're reported findings, not our data — but they're the kind of evidence a regulator points to when it wants to justify moving fast instead of studying the problem for another year.
Why this is worth an agent builder's attention
This story reads like a consumer-AI content story, but it's actually a preview of something builders in the onchain agent economy will hit eventually: regulators drawing a hard line around what an agent is *allowed to claim to be*, not just what it's allowed to do. China's rule is about emotional simulation specifically, but the underlying mechanism — a government deciding an agent's *presented identity* needs disclosure or restriction — is the same instinct behind every push for agent identity standards, onchain or off.
That's the connective tissue to the evidence-over-vibes approach this space needs generally: an agent's claimed persona, claimed safety, claimed trustworthiness is not the same as a verifiable one. China is drawing that line by banning a category outright. Onchain identity and reputation standards like ERC-8004 are trying to draw a version of that line by making an agent's identity and track record checkable instead of asserted. Different tools, same underlying problem — self-reported is not verified, whether the claim is "I'm your girlfriend" or "I'm a trustworthy trading agent."
What to watch
Watch whether Doubao and Qwen's task-oriented agent features (the non-companion functionality) survive the July 15 cutoff intact, and whether other Chinese platforms with similar companion features — there are several — follow with their own wind-downs before the deadline. Also worth tracking: whether this becomes a template other jurisdictions borrow from, since "restrict anthropomorphic AI, permit functional AI" is a cleaner regulatory line than most governments have managed to draw so far.
Sources
- ▸[ByteDance and Alibaba to Pull Agent Features as China Cracks Down on Humanlike AI — Decrypt](https://decrypt.co/372873/bytedance-alibaba-agent-features-china-cracks-down-humanlike-ai)