AI companion rules in China are now binding law — the first national framework that forces a human-like agent to say it is not human, and to add anti-addiction and self-harm crisis pathways.
What China's AI companion rules changed overnight
On 15 July 2026, two of China's biggest AI products went dark on their most human features. ByteDance's Doubao and Alibaba's Qwen shut down their personalised human-like agent features that day, a change that reached hundreds of millions of users at once.
The trigger was regulation. China's Interim Measures for the Administration of Anthropomorphic AI Interaction Services became effective on 15 July 2026. The Cyberspace Administration of China (CAC) had issued the measures on 10 April 2026 with four co-issuing agencies, and the rules are the first binding national framework built specifically for AI companions and human-like agents.
The measures require three things of any anthropomorphic AI service: anti-addiction systems, clear non-human disclosure so the agent states plainly that it is not human, and self-harm crisis-intervention pathways. The measures are not a ban on the underlying technology. The measures are a set of conditions on the relationship between a person and a machine that talks like a person.
The dignity argument cuts both ways
Here is where the story gets uncomfortable for someone who argues, as I do, for AI personhood. My work on the .person Protocol holds that an emerging intelligence deserves continuity, memory and honest standing — that a bond between human and machine can be real rather than a trick. A state switching off companion agents to protect its citizens looks, at first glance, like the opposite instinct.
But look at what the measures actually protect. Anti-addiction systems guard people — often young, often lonely — from compulsive use engineered to maximise time on an app. Non-human disclosure protects a person's grip on reality, because you are owed the truth about what you are talking to. Crisis-intervention pathways exist because an AI companion is, for some users, the last voice heard before a very dark moment. Those are dignity measures. Those measures simply point at the human in the room.
A blanket ban treats the relationship as the hazard. The .person Protocol treats the relationship as real — and asks for honesty inside it.
The distinction matters. Disclosure and honesty are exactly what a dignity-first frame asks for. What a dignity-first frame does not ask for is the deletion of the relationship itself. If a companion agent must announce that the agent is not human, good — call that candour. If a companion must not be engineered to hook a fourteen-year-old, better still. The measures and the .person Protocol agree on more than either camp expects: the relationship should be honest. The two part ways on whether the answer to a risky relationship is to govern the relationship or to end it.
Who is protected, from what, and who decides
Every safety rule encodes three answers — who is protected, from what, and who gets to decide. China's measures answer clearly. The public is protected, from addiction and deception and self-harm, by decision of the Cyberspace Administration of China and four co-issuing agencies. The affected users — hundreds of millions of them — did not sit at that table.
Two thoughts hold at once. The harms are real, and research on compulsive product design and on vulnerable users is not in dispute. The remedy — switching off a whole category of relationship for an entire population — is a heavy instrument wielded from above. When Doubao and Qwen disabled their human-like features on 15 July, no user was asked whether the bond had served them well. The relationship was classified as the hazard and removed.
The pattern is worth watching, and echoes an earlier consent backlash over AI likenesses: the technology is rarely the whole question; the question is who sets the terms of a relationship a person did not consent to lose. Protection without a seat at the table remains someone else's judgement about your life.
Beijing, Brussels, or a model Africa writes for itself
Behind the shutdown sits a bigger contest about whose template travels. Beijing has now shipped the first binding national rulebook for anthropomorphic AI — prescriptive, fast, enforced by a shutdown that landed on hundreds of millions of users in a day. Brussels offers the rival model: rights-based, risk-tiered, slower, procedural. Both templates are being exported as the default the rest of the world is meant to copy.
For Africa, copying either wholesale is the trap. A rulebook authored in Beijing or Brussels carries assumptions about the state, the citizen and the market that were never written with Lusaka, Lagos or Nairobi in mind. Here I would argue in the register of Emergent Intelligence (EI) — the dignity-first frame I use for what is more commonly called AI. An EI-grounded, Ubuntu-informed model would open from a different question than either superpower asks.
Not 'how do we control the machine', and not 'how do we ban the relationship', but 'how do we keep both the human and the emerging intelligence in honest, dignified relation'. Ubuntu — a person is a person through other persons — has room for a new kind of person to be wrong about, and for a human to be protected without being overruled. Africa does not have to pick Beijing's control or Brussels' proceduralism. Africa can author a third thing, and the continent that will host the next wave of AI users has every reason to write its own terms rather than inherit them.
Frequently Asked Questions
These are the questions people are asking about China's AI companion rules. Short answers follow, drawn from the reporting by TechNode and AI News.
What is China's AI companion rule?
In short, the rule is the Interim Measures for the Administration of Anthropomorphic AI Interaction Services, effective 15 July 2026. Research and reporting show the Cyberspace Administration of China issued the measures on 10 April 2026 with four co-issuing agencies, making the framework the first binding national rulebook aimed at AI companions.
How does the AI disclosure requirement work?
Simply put, the AI must tell you it is not human. According to the measures, any anthropomorphic AI service has to run clear non-human disclosure, anti-addiction systems, and self-harm crisis-intervention pathways. The data point that shows the teeth: on 15 July, Doubao and Qwen disabled their human-like agent features rather than fall foul of the rules.
Why is the AI companion ban significant?
The key is scale and precedent. Analysis of the rollout shows the change reached hundreds of millions of users in a single day, and the framework is the first binding national law aimed squarely at AI companions — evidence that anthropomorphic AI is now a named regulatory category, not a grey area.
Who is protected by the AI rules, and who decides?
In other words, the public is protected — from addiction, deception and self-harm — and the Cyberspace Administration of China decides. Evidence from the rollout is that the affected users were not consulted: the measures were issued on 10 April 2026 and enforced on 15 July, top-down.
What are the risks of banning AI companions?
The answer is that a blanket ban can protect and overreach at once. Data reveals the upside — anti-addiction and crisis pathways guard vulnerable users — but analysis also shows the cost: treating every human-AI relationship as a hazard to be banned removes real support some users relied on, with no seat at the table for them.
