China Switches Off Its AI Companions

The rules landed on 15 July, and the companions went quiet. China's Interim Measures for the Administration of Anthropomorphic AI Interaction Services are now in force, and the two largest consumer AI products in the country have responded by switching off the features that made them feel like friends.

ByteDance's Doubao and Alibaba's Qwen have both retired their personalised companion agents. We reported earlier this month that both companies had begun disabling these features ahead of the deadline. That preparation has now turned into a full shutdown. Doubao users have until 15 October to export their conversation histories. Qwen, according to reporting from Tech Times, is offering no migration path at all.

What the measures actually cover

The scope is narrower than the headlines suggest. According to the International Association of Privacy Professionals, the measures apply to services that simulate the personality traits, thinking patterns and communication styles of a real person. Work assistants are exempt. So are tools used in education and scientific research.

Providers who do fall inside the scope carry a specific list of obligations. They must be able to detect emotional distress in a conversation and route the user toward crisis intervention. They must build in measures against addictive use. They are barred from using sensitive personal data drawn from those interactions to train their models. The Cyberspace Administration of China supervises all of it, with fines up to 100,000 yuan for a standard breach and up to 200,000 yuan where a violation causes harm to someone's life, health or safety.

Those figures are small for companies of ByteDance's size. The compliance burden is the real cost, and evidently both firms judged the burden higher than the revenue.

A second framework for agents

The same date brought a separate piece of Chinese policy into effect. The Implementation Opinions on Intelligent Agent Governance is the first regulatory framework anywhere written specifically for AI agents rather than for models or chatbots.

Its central idea is a three tier authorisation structure. Agent actions are sorted by how consequential they are, and the threshold for requiring a human sign-off scales with that classification. Organisations deploying agents in high-risk sectors must file formally with regulators before they do so.

The pairing is instructive. Beijing has drawn a line between the agent that does your work and the agent that keeps you company, and it is the second kind that has been effectively legislated out of the consumer market. The first has been handed a permission structure and told to file paperwork.

Why this matters beyond China

Companion AI has been growing fast in every market, and the questions Beijing has answered are the ones regulators elsewhere are still circling. Who is responsible when a user in distress confides in a product rather than a person? Should a system that mimics intimacy be allowed to train on what it hears?

China has now answered both, bluntly, and the answers cost hundreds of millions of users their chat partners. Whether the trade was worth it is a question the rest of the world will get to consider with the benefit of watching.