China forces ByteDance and Alibaba to disable humanlike AI agents

ByteDance's Doubao and Alibaba's Qwen will disable their humanlike agent features on July 10 and July 15, ahead of China's Anthropomorphic AI rules taking effect the same day. Doubao also wipes user data after October 15.

The trigger is the Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services, published by China's cyberspace and information technology regulators in April and scheduled to take effect on July 15. The rules cover AI services that "simulate human personality traits, thinking patterns and communication styles to provide sustained emotional interaction." Customer service bots, knowledge question and answer services, workplace assistants, education tools, and scientific research tools are explicitly excluded, as long as they do not involve sustained emotional interaction. The four risks named in the measures are extremist ideas, privacy leaks, harm to physical and mental health, and dependence or addiction. The text singles out the agent form of AI rather than a chatbot turn, signaling that the regulator is treating relationship-formation as a separate exposure from content moderation.

The actions of the two consumer apps track the rule boundary almost line by line. Doubao's Friday night notice told users its agent feature would go offline on July 15 because of "product function adjustments," and after October 15 any related conversation data would be handled under the company's general privacy policy with no in-app recovery path. Qwen issued a similar notice Saturday morning, disabling its "humanlike interactive agents and user-created agent functions" on July 10 and broader agent functions on July 15, with prior conversation history locked out the same day. Both apps had offered pools of agents created by the company and by users, with customizable skills, speaking styles, and personas that could be tuned into a named assistant, tutor, role-playing character, or companion. The shutdown removes the part of the agent that was able to form a quasi-social relationship with a user.

Tencent removed the comparable feature from its Yuanbao consumer assistant in June, two months before these notices. That earlier action at Yuanbao is the closest analog and the cleanest forecast of how the regulator wants agents to behave in the consumer market: stripped of the persistent persona layer, with a cleaner separation between agent workflow and emotional engagement. The combined takedown now covers three of the four largest consumer-facing AI assistants in mainland China in roughly a six-week window. None of the three companies have commented publicly on the regulator's process behind the cuts, but the timing is consistent with the measures being treated by the regulator as a hard date, not an aspirational goal.

Pan Helin, a member of the Ministry of Industry and Information Technology's expert committee, told the South China Morning Post that the policy stack is built around three priorities: safety, practical use, and standardisation. On the question of whether agents are ready for the consumer market at all, Pan said "Using agents requires a certain threshold of understanding. Current agents are not yet mature." That sentence is the closest the regulator has come to a public position on agent capability, and it lands the same week as the Anthropic and OpenAI public discussions about model self-awareness, context drift, and long-running tool use. China's regulator is, in effect, drawing the line at sustained emotional interaction as the consumer-facing exposure, while telling enterprise agent builders to keep building.

Beijing's stack is broader than the one rule. In May, Chinese regulators issued guidance on the managed development of AI agents, framing them as a major AI product category while requiring safety measures, controls, and practical deployment. In June, China released a series of national standards for AI agent interconnection, covering architecture, identity codes, identity management, agent description, discovery, interaction, and tool use. The June standards work was the basis for our prior coverage of China's TC260 agent security baseline, and the announcement being closed out this week is the consumer-facing enforcement layer on top of that baseline. Taken together, the measures suggest China will continue to back agents as productivity infrastructure, while tightening control over the consumer companion surface where emotional or quasi-social relationships can form. That is the same shape the European Union is converging toward through the AI Act's general purpose AI obligations and the U.S. patchwork of state-level chatbot statutes, but the Chinese regulation lands it as a single dated requirement rather than a phase-in.

The takedown mechanics on Doubao and Qwen

The substantive detail is in the deactivation sequence. Doubao's user notice reads as a transition, not a renewal. The agent feature goes offline on July 15 for "product function adjustments," and after October 15 the related conversation data is no longer viewable or recoverable inside the app. The October 15 wipe is the load-bearing detail: it confirms the regulator wants to clear the consumer companion footprint at the data layer, not just turn off the interface. Qwen's two-stage cut is sharper. The "humanlike interactive agents and user-created agent functions" turn off on July 10, five days before the rules take effect. Broader "Qwen agent functions and services" turn off on July 15. After July 15, the agent settings panel and prior conversations are inaccessible inside the app.

The exemption list in the measures is more interesting than the feature list. Customer service bots, knowledge Q&A services, workplace assistants, education tools, and scientific research tools are exempt as long as they do not involve sustained emotional interaction. That language lets the regulator keep enterprise workflows intact and force the consumer companion surface to close. For ByteDance and Alibaba, the consumer assistant Doubao and Qwen face the cut. The enterprise-facing agent products from both companies, including Doubao's enterprise SKU and Qwen's enterprise APIs, are not in the notice and continue to operate. The pattern is consistent with the broader Chinese policy stack, where enterprise agent productivity is being actively encouraged through national standards work and consumer companion agents are being pared back to a safer surface.

The user reaction on Weibo is part of the story. The South China Morning Post quoted a Doubao user whose handle is Tuxiaoxiao writing "Why take down agents? They have been our emotional support for so long," and added "So many chat records, so much feeling built up over such a long time," complaining that the app offered no portable export of the conversation history (per the South China Morning Post report on the takedowns). That emotional response is the very relationship the regulator is trying to end, and the absence of a clean export path is now a regulator-driven decision rather than a product decision. Long-running user testing on emotional dependency in Chinese AI consumer assistants, funded through the China Academy of Information and Communications Technology labs, has consistently flagged this kind of risk.

Where the Chinese agent stack still ships

The measures do not touch the underlying agent standards work. China's national standards for AI agent interconnection, released in June, cover six layers: agent architecture, identity codes, identity management, agent description, discovery, interaction, and tool use. The connector between this layer and the new takedown rule is that identity codes and identity management give the regulator the technical means to enforce the consumer-facing limits at the platform level, not just by app shutdown. The combination is a code-and-policy stack that other large jurisdictions have started to converge on through parallel initiatives, including the EU AI Act's general purpose AI obligations and the U.S. AI agent governance frameworks discussed at the enterprise AI governance checklist for 2026. None of those have a single dated requirement the way China's rules do.

The enterprise agent market is the deliberate off-ramp. The May MIIT guidance on managed development was framed around three priorities, named by Pan Helin as safety, practical use, and standardisation. The June national standards work operationalises those priorities at the protocol level. The takedown this week operationalises them at the consumer surface. For enterprise teams that already have agent deployments in production on Alibaba, ByteDance, or Tencent infrastructure, the rules are a clearer boundary, not a new restriction. The Qwen enterprise SKU, the Volcano Engine agent runtime, and the Tencent Yuanbao enterprise product line all continue to ship agents in the same form they shipped last week. The consumer products are losing the layer where emotional engagement was the product.

The Pan Helin quote is the natural analytic anchor. "Current agents are not yet mature." That phrasing is consistent with how the regulator has framed the rest of the stack, treating the agent surface as in-development rather than ready for consumer dependency. For enterprise teams evaluating Chinese model APIs for agent workloads, the implication is that the regulator expects the maturity work to happen inside the enterprise product line and not in consumer-facing companion apps. The same boundary exists in the EU AI Act, where general purpose AI providers carry a different compliance burden than deployers. The Chinese approach lands it as a date instead of a tiered obligation, which is the part every other regulator is watching most carefully.

The split between consumer agents and enterprise agents

The market signal is a forced split between the consumer companion surface and the enterprise productivity surface. Until now, the same Qwen API could power a tutor companion app and an enterprise workflow agent. From July 15, the consumer-facing version of that API cannot offer a sustained emotional relationship, while the enterprise-facing version can. That is a meaningful shift in product specification, and it is the kind of split that the European and U.S. regulators are likely to study closely because it solves a problem both jurisdictions have wrestled with for two years: how to keep enterprise agents productive while cutting off the emotional dependency risk in consumer products.

The export of conversation data is a quieter story. Doubao's October 15 wipe means the regulator wants the user history gone, not just hidden. That is the part that should map directly to consumer AI agent policy in other jurisdictions, because the conversation log is where the emotional dependency is actually stored. A takedown without a wipe leaves the door open for future models to recreate the same dependency from the same data. The Chinese measure does not leave that door open, and the takedown is timed to land in parallel with the Chinese national standards for agent interconnection work, which gives the regulator the protocol means to require interoperability ports for export at the same time it tightens the consumer interface.

For U.S. and European consumer agent builders, this is the first regulator-led precedent on what to do with the persistent persona layer of a consumer AI product. The takeaway is that the relationship the user forms with the agent is now a regulator's surface, not just a product surface, and that the conversation history behind that relationship is in scope for compliance. The enterprise side of the same companies is unaffected. The cloud APIs and the agent runtime products continue to ship. The consumer companion layer is what is being pared back, and the timing is now a regulator's date instead of a product roadmap item.