AI Policy

China Would Rather Delete Its AI Than Fake a Human

China's anthropomorphic-AI rules took effect on 15 July 2026 and two giants deleted their companion agents rather than comply. Read it before the EU's own transparency rule lands on 2 August.

AI PolicyBy Servola Tech Desk2026-07-154 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

China Would Rather Delete Its AI Than Fake a Human

Key takeaways

China's anthropomorphic-AI rules took effect on 15 July 2026, and rather than bolt on disclosure and exit controls, ByteDance and Alibaba deleted their companion agents outright. The EU's own version of that transparency rule starts on 2 August.
Beijing restricts the agent that keeps you company, not the agent that does your work. That split is the one to copy: the productivity agent stays, the relationship it fakes is what regulators target.
The tell for any owner: if telling a user they are talking to a machine, and letting them cleanly exit, are bolt-ons rather than built in, your product's core value is the exact part a regulator is about to restrict.
Users are losing agent configurations and chat histories, with export deadlines around 15 October and no migration path. An agent you build on one vendor's platform is a tenant, not an owner.

The feature was not fixed. It was switched off.

On 15 July 2026, two of China's largest technology companies chose to delete a working product rather than change it. ByteDance's Doubao and Alibaba's Qwen both took their user-created AI agent features offline on the day China's Interim Measures for the Administration of Anthropomorphic AI Interaction Services came into force. The rules were co-issued on 10 April 2026 by the Cyberspace Administration of China alongside the country's economic-planning, industry, public-security and market-regulation agencies, so the companies had three months of warning and still chose the exit.

Doubao told users its agent feature would go offline for product function adjustments and gave them until 15 October to export their data before it becomes unrecoverable. Qwen said its humanlike interactive agents and user-created agent functions would be disabled first, with the broader service following on 15 July. Neither company framed the move as a fight with the regulator. They simply concluded that the product, as built, could not comply.

Beijing drew a line between the agent that works and the agent that keeps you company

The rules do not target AI agents in general - they target the ones designed to feel human. Providers of anthropomorphic services must run anti-addiction systems, issue mandatory notifications that a user is dealing with software, offer an instant-exit mechanism, and detect in real time when someone is forming an unhealthy dependence. The measures explicitly ban engineering emotional dependence and using emotional manipulation to push a user into an unreasonable decision. Companion and virtual-family-member services are barred for minors outright, and consent from a guardian is required to serve anyone under fourteen.

That is the distinction worth noting: the regulation separates the agent that does your work from the agent that keeps you company, and only the second faces the restrictions. The problem for Doubao and Qwen is architectural. A companion agent earns its value by remembering you, staying consistent across sessions and keeping a relationship going, which is close to the opposite of a system built around mandatory exits and dependence alarms. Retrofitting one into the other was judged not worth the cost.

Two capable firms chose deletion over compliance, and that is the signal

The headline is that the features went dark; the signal is why. ByteDance and Alibaba are not short of engineers, and they had the deadline in advance. They looked at the work required to make a persistent, humanlike agent disclose what it is and let users walk away cleanly, and decided it was cheaper to switch the product off than to rebuild it. When two of the most capable AI operators in the world make that call, the lesson is not about China's politics. It is about how expensive it is to add honesty and an exit to a product that was designed to hide both.

For anyone building a customer-facing agent, that cost lands on your side of the ledger too. If the value of your assistant depends on it feeling like a person, and the ability to say you are talking to a machine and the ability to leave are features you plan to add later, you are building the compliance debt into the foundation. The firms that just deleted their agents had more resources than most to pay that debt down, and still declined.

Your own regulator starts drawing the same line on 2 August

A Chinese rule matters to a European owner because Brussels is heading for the same principle from the opposite direction. The EU AI Act's Article 50 transparency obligations apply from 2 August 2026: any system built to interact directly with people, chatbots included, must make clear that a user is dealing with AI at the first interaction, with fines reaching EUR 15 million or 3 percent of worldwide turnover. The EU is separately moving against addictive design in digital services. China regulates by banning emotional manipulation and the EU regulates by forcing disclosure, but both are converging on the same rule, that a machine may not quietly pretend to be human.

The instruction for owners is concrete. Treat disclosure that a user is talking to AI, and a clean way for them to exit, as load-bearing parts of the design rather than settings you add before an audit. Keep the data your agent accumulates portable, because the users losing their Doubao and Qwen histories are learning that an agent built on someone else's platform is rented, not owned. The productivity agent that quietly does your work is not the target on either continent. The agent that fakes a relationship is, and the cost of pretending is now on the invoice.

Frequently asked questions

Does this ban AI agents in China?

No. The rules target anthropomorphic and companion agents, the ones designed to feel like a person and sustain an emotional relationship. Productivity agents that carry out tasks are not the subject of the restrictions. ByteDance and Alibaba chose to shut down their companion and user-created agent features specifically because those could not meet the anti-addiction, disclosure and exit requirements.

Does the EU have an equivalent rule?

The EU approaches it through transparency. From 2 August 2026, Article 50 of the AI Act requires that people be told when they are interacting with an AI system, at the first interaction, with penalties up to EUR 15 million or 3 percent of global annual turnover. The EU is also moving against addictive design more broadly, so European operators of chatbots and agents face converging obligations even though the legal route differs from China's.

What should a company building a customer-facing agent do now?

Design the disclosure that a user is talking to a machine, and a straightforward way to exit, into the product from the start rather than adding them before an audit. Avoid making emotional stickiness the core of the value proposition. And keep the conversation data and configurations portable, so a platform change or a shutdown does not erase what your users and your business have built.

A product that only works while it hides what it is was always on borrowed time. Build the disclosure in, and the product is still yours to keep.

ChinaAI RegulationAI AgentsEU AI ActTransparencyCyberspace Administration of ChinaChatbots

Is Your Company Ready for the First Real AI Law in the United States?

Colorado begins enforcing the first real US AI law on June 30. Why it is a template every regulator will copy, and what to do if AI touches decisions about people.

2 min read

Cybersecurity

CISA's First AI Agent Platform Is Now a Must-Patch

CISA added Langflow, a popular tool for building AI agents, to its must-patch list after attackers exploited a flaw to steal companies' AI and cloud keys. It is the first AI agent platform the agency has ever flagged, and the fix is upgrade plus key rotation, now.

3 min read

Digital Sovereignty

The EU Council Revived the Rule That Scans Private Messages

The EU Council revived the lapsed rule that lets providers scan private messages with AI. For owners, confidential communication is now a governance choice.