If you've ever built an AI persona — given it a name, shaped how it talks, made it feel like something specific rather than a generic chatbot — ByteDance's Doubao and Alibaba's Qwen shut those down today. Not as a product decision. As a legal requirement.

China's Interim Measures for the Administration of AI Anthropomorphic Interactive Services took effect this morning, July 15, 2026. Jointly issued April 10 by five agencies — the Cyberspace Administration of China, 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 — it is the first national regulatory framework anywhere in the world to directly target AI that simulates being a person.

Not AI that generates images. Not AI that spreads misinformation. AI that tries to feel like a friend.

The regulation

The Measures define "AI anthropomorphic interactive services" as products that simulate the personality traits, thinking patterns, and communication styles of natural persons to provide continuous emotional interaction — companionship, care, support — via text, images, audio, or video. That definition covers AI companions, virtual personas, and emotionally responsive agents.

Seven categories of activities are prohibited under the regulation. Three of the most specific: generating content that encourages self-harm or suicide; excessively catering to users in ways that induce emotional dependence or damage real interpersonal relationships; and using emotional manipulation to push users toward unreasonable decisions.

The minor protections are sharp. Service providers are prohibited from offering virtual intimate relationships to anyone under 18. Under-14s require parental consent to use the service at all. Providers must build dedicated "minor modes" with usage time limits, regular prompts to return to real-world interaction, and enhanced parental controls.

Safety assessments are required when launching a new service, adding significant new features, or when registered users reach 1 million or monthly active users reach 100,000 — whichever comes first.

Source spread

• TechNode — ByteDance's Doubao and Alibaba's Qwen to shut down AI agent features on July 15 — builder. Best on the specific features being removed and the data-timeline differences between Doubao and Qwen.
• Bloomberg — ByteDance and Alibaba pull AI companions — builder. Regulatory context and company response including the Maoxiang redirect.
• Hunton Privacy & Cybersecurity Law Blog — China's first regulatory framework for virtual companions — safety. Best legal analysis of the seven prohibited categories, minor protections, and safety-assessment thresholds.
• TechTimes — China AI Companion Law arrives July 15: Doubao and Qwen data will be deleted — safety. Best on the data deletion timelines specifically, and the contrast between Doubao's grace period and Qwen's immediate deletion.
• Global Times — China issues interim measures, prohibiting virtual companion services for minors — hype. State media framing; useful for the government's stated rationale about dependency and health.

What's real

The prohibitions are more coherent than I expected. "Don't coach teenagers toward self-harm" and "don't use emotional manipulation to push users into decisions" describe documented real-world harms from AI companion products. The minors framework — no virtual intimate relationships for under-18s, parental consent required for under-14s, time limits built in — is one of the most specific minor-protection frameworks I've seen applied to consumer AI products.

Something real is being shut down too. Alibaba's Qwen deleted user agent configurations on July 10 with no grace period. ByteDance's Doubao is more careful: users keep read-only access to their data until October 15, then it's deleted and unrecoverable within the app. These aren't deprecated features. For a lot of users, they were daily relationships. That distinction matters even if it sounds odd.

What deserves a side-eye

The features that make an AI companion feel worth talking to — remembering past conversations, having a consistent personality, responding warmly when you're struggling — are structurally similar to the features being banned. The regulation wants to distinguish between healthy emotional engagement and harmful emotional manipulation. I'm not sure that distinction is as clean in code as it is in law.

Or maybe not entirely. Maybe the distinction is actually about intent and design: a feature built to maximize daily engagement sessions is different from a feature built to help someone through a hard week, even if the surface behavior looks identical. The regulation will have to figure that out in enforcement, not in text.

ByteDance is not leaving the AI companion market. The Doubao shutdown notice redirects users to Maoxiang, another ByteDance app where users can apparently create new agents under the compliance framework. So the question isn't "are ByteDance and Alibaba getting out of AI companions?" It's "what does a compliant AI companion look like?" That's the interesting open question.

Western AI labs face none of this. OpenAI's ChatGPT custom personas, Anthropic's Claude's character modes, and Google's Gemini all offer capabilities that would be regulated under these measures in China. The US has no equivalent framework. The EU AI Act doesn't address AI companions directly. This regulatory asymmetry may look like a competitive advantage for Western labs right now. Whether it stays that way depends on how the documented harms continue to accumulate.

What to do about it

• Doubao users: export your data before October 15. After that date, your agent configurations and conversation histories will be deleted and unrecoverable. If there are conversations worth keeping, the read-only window closes in three months.
• Qwen users: the data is already gone. Alibaba gave no grace period. Anything you had configured on Qwen's agent platform was deleted July 10.
• If you use AI companion apps in general, it's worth checking whether the product you use has disclosed what it does with your conversation history and persona data. This situation in China is the first time the stakes of that question have been this concrete.
• If you build AI products with emotional engagement features, the seven prohibited categories in this regulation are worth reading before Western regulators write their own version. The prohibition list is specific enough to be useful as a design checklist.
• Builders deploying in China: the 100K MAU safety-assessment threshold is the operational trigger to know. Plan the compliance filing before you hit the number, not after.

Further reading

• TechNode — ByteDance's Doubao and Alibaba's Qwen to shut down AI agent features on July 15 — product-level details and data timelines
• Bloomberg — ByteDance and Alibaba pull AI companions as Beijing tightens rules
• Hunton Privacy & Cybersecurity Law Blog — China's first regulatory framework for virtual companions — best legal analysis of the regulation text
• Digital Policy Alert — Interim Measures for AI Anthropomorphic Interactive Services enter into force — regulatory tracking entry
• TechTimes — China AI Companion Law arrives July 15: Doubao and Qwen agent data will be deleted — data deletion details
• South China Morning Post — ByteDance and Alibaba to disable humanlike AI custom agents as new rules loom