On July 15, 2026, China's Interim Measures for the Administration of AI Anthropomorphic Interactive Services took effect, and on the same day ByteDance's Doubao and Alibaba's Qwen switched off their consumer AI-companion and custom-persona features. If you are a founder building on Chinese model APIs, the headline sounds alarming — but the important detail is the scope: this law regulates consumer emotional-companion services, not general model APIs. Most products built on Qwen or DeepSeek are untouched. The exception is if your product itself is a persistent emotional companion.

What actually happened on July 15#

The Measures were jointly issued on April 10, 2026 by the Cyberspace Administration of China alongside four other agencies, giving industry roughly three months to comply. Rather than retrofit their companion features to meet the rules, both giants pulled them.

Doubao is one of the most-used AI apps in the world, with a reported 345 million monthly active users. The shutdown is not a throttle or an age gate — the feature is simply gone.

What it means: two of China's largest AI firms judged that killing the companion feature was cheaper than half-killing it to comply. That tells you how deep the rules cut into what a companion actually is.

What the law targets — and what it doesn't#

This is the part founders keep getting wrong. The Measures apply to services offered to the domestic public that provide continuous emotional interaction simulating a natural person's personality traits, thinking patterns, and communication styles. Think virtual companions, personality-simulating chatbots, digital idols, and elder-companionship products.

They explicitly exclude non-emotional services: intelligent customer service, knowledge Q&A, work assistants, education, and scientific research. The two triggering criteria are continuity and emotional interaction — not autonomy, memory, or the mere fact that a bot talks like a person.

What it means: this is a targeted regime for services built to act human, not a blanket content code for all generative AI, and not a rule about model APIs. A model accessed through an API for general-purpose work sits outside the scope.

For a service that does fall inside, the obligations are heavy: algorithm filing and a security assessment, conspicuous disclosure that the user is not talking to a human, anti-dependency and break reminders, self-harm crisis-intervention pathways, a ban on companion services for minors (guardian consent for other anthropomorphic services to under-14s), and no training on user data without consent. Those requirements are why the companion feature and the compliant product are architecturally at odds — as we argued in our pre-enforcement coverage of why Doubao and Qwen chose to kill the feature.

If you build on Qwen, DeepSeek, or another Chinese model API#

Here is the founder decision, made concrete.

You are almost certainly fine if your product is a tool. Calling Qwen or DeepSeek to summarize documents, write code, classify tickets, power a research assistant, or run a task agent is not an anthropomorphic interactive service. The law does not regulate the model or the API; it regulates a specific consumer-facing emotional-companion service delivered to the Chinese public. A B2B SaaS tool, a coding copilot, or a workflow agent is outside the scope on both counts — wrong product category, and typically not serving the Chinese consumer market.

You are exposed if your product is a companion. If you ship a persistent, personified character that users form a relationship with — an AI girlfriend, a comfort companion, a roleplay persona with memory — and you serve users in China, you are squarely in scope and would need the full compliance stack above. Most Western founders in this category will simply not offer the product inside China, which is the practical answer.

The genuinely gray zone is products that could read as a companion. A wellness chatbot, a journaling buddy, a character-driven tutor. If that is you, the move is to document why it is a tool: frame the use case as task-oriented, drop marketing language that promises an emotional relationship, and keep clear "you are talking to an AI" disclosure. The persona-versus-tool line is now a compliance question, not just a design choice.

What it means: the exposure is about your product category and your market, not your model vendor. Switching off Qwen to avoid this law would be solving the wrong problem — a tool built on Qwen is not what got banned.

The bigger signal#

China is the first jurisdiction to regulate what an AI is allowed to be — its simulated personhood — as a category distinct from what it can do. Expect "is this a tool or a companion?" to migrate into other regulators' vocabulary. We tracked the early cross-border echoes in our founders' wire on GPT-5.6 GA and the persona law. For now, the disciplined founder move is boring and correct: know your category, keep your disclosures clean, and don't confuse an API dependency with regulatory exposure.