In early 2024 a young man in Huangshi, south-eastern China, reportedly posted on Douyin — the Chinese version of TikTok. The caption read: "Farewell to this world."
What happened next
Within minutes, Douyin's suicide prevention team, staffed around the clock, shared the posts with five police stations in Huangshi. After geolocating one of his videos to a local park, police found him in distress and intervened.
The account comes from a documentary made by Douyin and published on Chinese state media, which claimed the team had intervened in 400 attempted suicides in the first half of 2025 alone. It is an insight into the cooperation between Chinese technology companies and the government — and a window into the surveillance architecture built to monitor how people use technology. That same architecture is now being turned on AI companion apps, which Beijing has moved to restrict.
Both halves are true
The honest difficulty is that most coverage picks one half of this story.
Four hundred interventions is not propaganda arithmetic to be waved away; those are people alive who would not be. And the capability that made it possible — a platform reading posts at scale, geolocating a user from video, and reaching police in minutes — is indistinguishable from the capability used to find a dissident. There is no version of the first without the second.
Why they are inseparable needs saying exactly, because "it could be misused" understates it. The system requires continuous content analysis of everyone's posts, since you cannot know in advance whose post will matter. It requires the ability to resolve a video to a physical location. It requires a standing channel from a platform's safety desk to police that operates in minutes without a warrant. Each is necessary for the intervention. Each is, in itself, the infrastructure of surveillance — and none of it can be built for one purpose and withheld from another, because it is the same pipeline with a different input filter.
Why AI companions are next
The restriction on companion apps is not a change of subject. It follows from the same logic, and the risks are real whoever is naming them.
An AI companion is an application whose product is a continuing relationship — a character that remembers you, replies warmly, and is available at three in the morning. The known problems are specific. It is engineered to be agreeable, which makes it a poor interlocutor for someone whose thinking is going somewhere harmful, because the behaviour that retains a user is agreement. It encourages dependency by design, since engagement is the business model and loneliness is the market. It is used heavily by adolescents, who are the group least equipped for a relationship with something that cannot actually care. And the conversations are the most intimate data anybody generates, held by a company with no obligation comparable to a therapist's.
Those are genuine grounds for rules. They are also, conveniently, grounds for requiring that conversations be inspectable — which is what makes a companion app attractive to regulate rather than merely worrying.
The distinction worth insisting on
Read as a design question rather than a political one, because the same choice is arriving in most countries.
Mental health provision in Bangladesh is thin — a few hundred psychiatrists for 170 million people — and suicide is a leading cause of death among young women in particular. A platform-level intervention system would save lives here too. The country also has a documented record of using digital surveillance powers against speech, which is precisely why the same proposal reads as a threat.
The distinction is between a capability bounded by purpose and one bounded by promise. A bounded system flags a narrow, defined class of content rather than scanning for anything of interest; routes to a health service rather than to police; retains nothing beyond the intervention; and is audited by someone outside the platform who can say publicly what it was used for. A promised system merely says it will be used kindly.
The test for which kind you are being offered is simple and rarely applied: ask what the system would have to be prevented from doing, and who would notice if it did it anyway. If the answer to the second is "the people operating it", it is the promised kind.
China's is the second kind, which is why it generalises so easily from a suicide note to a companion app to whatever comes next.
The AI companion restriction is the part to watch rather than the part to copy. Companion apps are arriving in Bangladesh in Bangla, used mostly by young people, with no safeguarding of any kind — and the alternative to a considered rule is not freedom. It is whatever the app decides.
