Three months. That is how long it took from April 10, 2026, when the Cyberspace Administration of China and four other ministries jointly announced the Interim Measures for the Administration of Anthropomorphic AI Interaction Services, to their enforcement on July 15. On that enforcement date, ByteDance's Doubao and Alibaba's Qwen are switching off their personalized AI agent features. Doubao cited product adjustments as the reason.
The term anthropomorphic is not complicated. It refers to services that imitate a person's personality, speech and emotions and keep exchanging feelings with the user. Virtual lovers, virtual family members and emotional companion AI fall under it.
The interim measures require platforms running such services to strengthen content review, protect minors and put in place mechanisms that prevent excessive emotional dependence. Content that encourages or glamorizes self-harm or suicide, content harmful to minors, and practices that induce user dependence are also listed as prohibited items.
Why the companies switching off these features chose termination over modification shows in the design. Personalized agents are built on the premise of remembering the user and maintaining a relationship. That is fundamentally at odds with requirements to cut off conversations forcibly at a certain point and to keep a distance so users do not grow attached. Qwen issued a phased notice saying it would take down humanlike agents and user-created agents on July 10, and clear out the rest on the 15th.
Doubao has been expanding personalized agents since 2024, and Qwen has run a service where users build and operate their own characters. These are features accumulated over more than two years. Doubao users can still view and download agent information and conversation records for a while after the shutdown, and this data will all be deleted on October 15.
Enforcement was under way even before the effective date. The Shanghai Municipal Cyberspace Administration said on June 26 that it had deleted more than 14,000 AI agents that impersonated government agencies, encouraged gambling or collected personal information without authorization.
China is not pushing out agents themselves. An implementation guideline supporting development, issued in May by the Cyberspace Administration of China together with the National Development and Reform Commission and the Ministry of Industry and Information Technology, defined an AI agent as an intelligent system that perceives, decides and interacts on its own. The same guideline set down in writing for the first time the principle that users can know about and refuse an AI's decisions all the way through.
Consumer agents have carried the assessment that they are used frequently but are hard to monetize because server and computing costs are heavy. Qwen recently opened an agent service aimed at businesses and developers rather than general users, with Luckin Coffee, KFC and China Eastern Airlines taking part in testing. A forecast that the enterprise AI agent market will grow past 48 billion yuan, over 9 trillion won in Korean currency, and a Gartner projection that up to 40% of enterprise applications will include agents by 2026 are cited as the backdrop for this shift.
A supply chain issue has been added on top. Alibaba, citing the risk of backdoors being planted in Anthropic's Claude Code, decided to have all employees delete Anthropic models and agent products starting July 10 and to move internal AI work to its own agent, Qoder. An engineer on Anthropic's side explained, regarding suspicions of collecting Chinese user information, that it was an experimental mechanism intended to block unauthorized third-party circumvention access and model theft, and that it would be removed. Anthropic has also claimed that Alibaba and operators connected to Qwen used its models through large numbers of fake accounts.
The first case has emerged in which a service's survival was decided by when to cut off a conversation and whom to screen out, more than by how much performance was raised. Any company in Korea preparing an emotional companion service should draw the shutdown button and minor verification into the first page of the blueprint instead of pushing them to the back of the feature list.
On the morning of July 15, someone will open a chat window they have been talking to for nearly two years and see a screen with no reply coming back. The conversation records disappear three months later.
