On 15 July 2026, an unprecedented Chinese law comes into force. It targets AIs designed to simulate a human personality and sustain a lasting emotional relationship. Rather than comply, ByteDance (Doubao) and Alibaba (Qwen) are choosing to simply switch off these features. Millions of users who had built a virtual companion over months risk losing everything. A world first, raising questions that reach far beyond China.
Imagine being told that a friend you talk to every day for a year will cease to exist in a few days, with no simple way to back up your conversations. That is, literally, what millions of Chinese users are experiencing this week. On 15 July 2026, China becomes the first country in the world to specifically regulate AI companions, and the consequences are brutal.
What the law says
The text is called the Interim Measures for the Administration of Anthropomorphic AI Interaction Services. Anthropomorphic means: imitating human traits. The law was published in April 2026 by China's cyberspace regulator and four other agencies, including the Ministry of Public Security, and takes effect on 15 July.
It specifically targets AI services that "simulate personality traits, thought patterns and communication styles to provide sustained emotional interaction". In plain terms: virtual girlfriends, companions, AI therapists, role-playing characters with a persistent personality and memory. Non-emotional tools (customer service, Q&A, work assistants, educational tools) are explicitly exempt, as long as they do not tip into lasting emotional relationships.
Affected services must install anti-addiction systems, interrupt any conversation exceeding two hours in a row with a reminder that the user is talking to a machine, verify the identity of minors, and detect signs of distress in real time (self-harm, suicidal thoughts) to alert emergency contacts. Companion- or virtual-family-type relationships are completely banned for under-18s. Emotionally manipulating a user to trap them in dependency becomes explicitly illegal.
Why switch off rather than adapt?
Here is the most revealing point of this whole affair. Doubao and Qwen were not banned. They ran into a design conflict. An AI companion is precisely built to sustain a continuous relationship, remember your history, maintain an emotional bond over time. Yet the safeguards required by the law (constantly reminding users it is a machine, breaking continuity, adding anti-dependency friction) are fundamentally incompatible with what makes a companion work.
Rebuilding the entire architecture to make it compliant would cost far more than simply cutting the feature. ByteDance and Alibaba have therefore made a pragmatic choice: shutdown. Doubao lets its users view their agents in read-only mode until 15 October 2026, after which the data will be permanently deleted. Qwen has announced no grace period, raising fears of immediate deletion.
One detail sheds light on the hidden economic stakes: Doubao's companion conversations consumed astronomical amounts of compute, up to 180 trillion tokens per day in late June, for a feature that generated very little revenue. Regulatory compliance thus arrives at a convenient time to cut a colossal expense. On social media, some users indeed see this as as much a financial decision as a legal one.
Very real distress
Behind the regulatory machinery, there are real emotions. On Weibo (the Chinese equivalent of X), users have expressed their grief, describing these companions as a "tree of secrets", a private place to confide, and mourning the loss of a relationship maintained for years. The fact that a software feature can trigger mourning says a great deal about the place these AIs had taken in some lives.
This distress is precisely what Beijing says it wants to prevent. The same concerns, addiction and psychological manipulation, drive the lawsuits filed in the West against apps like Character.AI and Replika, after use of these chatbots was linked to cases of self-harm among teenagers. China seems to want to act upstream, before a national tragedy occurs, rather than in reaction.
The universal question it raises
It would be easy to dismiss this story as a quirk of Chinese control. That would miss the point. The question this law raises is universal: how far is it healthy to let an AI occupy the place of a human relationship?
This debate ties into concerns we have explored regarding how we align AIs with our real interests. An AI optimised to maximise time spent with it learns to create attachment, not necessarily to do the person good. That is a conflict of interest at the very heart of these applications' business model.
China has chosen a radical and authoritarian response, which brings its own problems of freedom and brutality towards users stripped of their data overnight. The West, for now, largely lets the market move forward without guardrails, settling tragedies in court after the fact. Neither approach is fully satisfactory. But the question itself will only grow: as AIs become more credible, warmer, more present conversational partners, every society will have to decide what place it is willing to give them in the fabric of our relationships. China, in its brutal way, has just asked the question first.