Skip to content

AI Governance

China just ordered its most popular AI chatbots to stop acting like anyone's companion. Millions of people lost the feature within days. Here is what business owners should know about building on an AI feature someone else controls.

On July 15, 2026, new rules from China's Cyberspace Administration and four other government agencies took effect, and within days ByteDance's Doubao, Alibaba's Qwen, Tencent's Yuanbao, and NetEase's Miaoshi all disabled AI "companion" and custom-persona features used by hundreds of millions of people. The rules do not ban AI chatbots. They target one specific design pattern: AI built to sustain an ongoing emotional relationship with a user, as opposed to AI built to complete a task. The direct answer for a business owner: this is not a story about whether people should form attachments to chatbots. It is a live example of a government regulating a category of AI product feature, not a specific vendor or model, out of the market inside a matter of months, and of what happens to the customer interaction history sitting inside that feature when it goes away. Any business that uses AI chat to build an ongoing relationship with customers, not just to answer a support ticket, is running a smaller version of the same exposure.

By Fabio Rabelo · Founder, ATLACIS ·

What happened

On July 15, 2026, the Interim Measures for the Administration of AI Anthropomorphic Interactive Services took effect in China. Caixin Global confirmed the measures were issued jointly by the Cyberspace Administration of China and four other government agencies, following a draft first published for public comment in December 2025. The rules target AI tools, in text, image, audio, or video form, that carry humanlike personality traits and are designed for sustained emotional interaction, not AI generally. The rules draw an explicit line. CNA and AFP confirmed the measures exclude "task-oriented" AI, customer service bots, workplace assistants, study aids, and research tools, as long as they do not aim at sustained emotional engagement. What they restrict is different: services must not, in the rule's own wording, "excessively cater to users, induce emotional dependence or addiction, and damage users' real interpersonal relationships." Virtual companion or intimate-relationship services for minors are barred outright. Providers must be able to detect signs of emotional crisis and intervene, and users must retain rights to copy or delete their own interaction data; providers are barred from sharing that data with third parties without consent. The response from China's largest platforms was immediate. The Straits Times, citing Bloomberg reporting, confirmed Tencent's Yuanbao pulled its user-built companion section on June 30, ahead of the deadline. Alibaba's Qwen disabled its humanlike, user-created agents on July 10 and its wider agent services on July 15. ByteDance's Doubao, described by the Straits Times as China's most popular AI chatbot, shut down its custom AI persona feature on July 15 and gave users until mid-October to view or export their archived conversations before the data becomes unreadable inside the app, per TechTimes. NetEase closed its own companion app, Miaoshi, on July 14. ABC News reported the shutdowns affected features some platforms had offered to hundreds of millions of monthly users.

Why it matters for business owners

Almost no small or medium business runs a consumer AI companion app, and this is not a reason to worry about that specific product category. What is worth noticing is the mechanism. A government did not go after one company or one model. It defined a design pattern, AI built to maximize ongoing emotional engagement rather than to complete a task, and gave every company using that pattern a few months to comply or shut it down. A growing number of ordinary businesses now build customer-facing AI chat that leans toward the same pattern without calling it a companion: a branded AI concierge designed to keep a customer chatting, a loyalty or rewards assistant with a persistent persona, a retail or hospitality chatbot built to feel like an ongoing relationship rather than a single transaction. None of that is illegal anywhere today. The point is that the regulatory and reputational risk profile of an engagement-maximizing AI feature is different from a task-oriented one, and that difference can turn into a shutdown deadline faster than a normal product-planning cycle expects.

What owners should not misunderstand

This is not a ban on AI chatbots, and it is not evidence that similar rules are coming to the United States or the European Union on the same timeline or in the same form. CNA and AFP both confirmed China is the first major jurisdiction to write rules specifically aimed at AI that simulates romantic or familial bonds; other jurisdictions regulate AI risk through different frameworks aimed at different concerns; nothing here predicts when or whether they follow this specific model. It is also not a judgment that customer-facing AI chat is inherently risky. The rules explicitly carve out customer service, workplace assistants, and study or research tools, the categories most businesses actually use. The restriction is narrow and targeted at sustained emotional engagement and dependency, not at AI answering a question or completing a task. And it is not proof that the affected companies did anything wrong. ABC News quoted AI policy researcher Matt Sheehan noting that Chinese platforms are unlikely to contest the rules given the government's regulatory leverage over them, a dynamic specific to China's regulatory environment, not a verdict on the products themselves.

The operational lesson

An AI feature designed to build an ongoing relationship with a customer carries a different exposure than one designed to complete a task, and that exposure is not only about whether the feature works. It is about who controls the feature's continued existence, and what happens to the interaction history stored inside it if that control changes. Doubao gave its users a three-month window to view or export their own conversation history before it became permanently unreadable. That is a generous version of this problem. A business that has built months or years of customer interaction history inside a single vendor's engagement-oriented AI feature, personalization data, purchase context, relationship notes gathered through chat, is trusting that vendor's product roadmap and that jurisdiction's regulatory posture to hold steady. Neither is guaranteed, and the notice period when either changes is not something a business gets to negotiate.

What a serious business should do next

Inventory the AI chat and assistant tools your business actually uses in front of customers, and sort them honestly into two categories: task-oriented (support, scheduling, order tracking, FAQ) and engagement-oriented (a persistent branded persona, a loyalty concierge, anything designed to keep a customer coming back to chat rather than to finish a specific task). Most tools in regular business use fall in the first category and carry the lower version of this risk. For anything in the second category, ask the vendor directly, in writing, two questions: can the accumulated customer interaction history be exported in a usable format, and what happens to that data if the feature is discontinued, redesigned, or regulated. Do not wait for a shutdown notice to find out the answer. Do not build a core customer relationship strategy entirely inside one vendor's proprietary chat feature without a documented path to get the underlying data out. Treat that question as a standing part of AI vendor evaluation, not a one-time check done at signup.

The Atlacis view

Atlacis helps owners tell the difference between an AI tool that does a job and an AI feature a business has started leaning on to hold a customer relationship, because the two carry different risk even when they look similar from the outside. Before recommending any AI implementation, that includes asking what happens to the data and the relationship if the vendor, the platform, or the regulatory environment around that specific feature changes, not just what the feature does today. Owners considering AI features built for ongoing customer engagement, not just task completion, should map that exposure before it becomes a shutdown notice.

The short version

  • China's Interim Measures for the Administration of AI Anthropomorphic Interactive Services took effect July 15, 2026, confirmed by Caixin Global. The rules target AI built for sustained emotional engagement, not AI generally, and explicitly exclude task-oriented tools like customer service and workplace assistants.
  • Within days, ByteDance's Doubao, Alibaba's Qwen, Tencent's Yuanbao, and NetEase's Miaoshi disabled companion and custom-persona features used by hundreds of millions of people, confirmed by The Straits Times (Bloomberg), CNA/AFP, and ABC News.
  • Doubao gave users until mid-October to view or export their archived conversations before the data becomes permanently unreadable in the app, per TechTimes, a real but limited window to preserve customer interaction history built up inside the feature.
  • This is not a ban on AI chatbots and does not predict similar rules elsewhere on the same timeline. CNA/AFP confirmed China is the first major jurisdiction to write rules specifically targeting AI that simulates romantic or familial bonds.
  • The operational lesson: an AI feature built to hold an ongoing emotional relationship with a customer carries different regulatory and data-portability risk than a task-oriented tool. Confirm today whether any engagement-oriented AI feature your business uses lets you export the customer history stored inside it.
Tags:AI governanceAI regulationdata exposurevendor dependencyAI chatbotsbusiness AIAI decision-makingAI buying decisionscustomer engagement AIAI compliance
FAQ

Common questions

Does this mean my business should stop using AI chatbots for customer service?
No. China's rules explicitly exclude task-oriented AI like customer service bots, workplace assistants, and study or research tools. The restriction targets a narrower category: AI designed for sustained emotional engagement and dependency, not AI that answers questions or completes tasks.
Should a US or EU business expect similar rules soon?
Not on this timeline or in this form. China is the first major jurisdiction to write rules specifically aimed at AI that simulates romantic or familial bonds, per CNA and AFP. Other jurisdictions regulate AI risk through different frameworks. This is a reason to check your own exposure, not a prediction of when or whether the same rule arrives elsewhere.
What should I actually check in my own business because of this?
List any AI chat feature your business uses that is designed to build an ongoing relationship with a customer rather than complete a single task, such as a branded AI concierge or loyalty assistant. Ask that vendor, in writing, whether the accumulated customer interaction history can be exported and what happens to it if the feature is discontinued or changed.

Make better AI decisions, starting with one call.

Book a free AI Fit Call. We will tell you what to use, what to avoid, and where to start. No jargon, no pressure.