Today, 15 July 2026, ByteDance takes Doubao’s companion-agent function offline across China, the visible edge of a rule that has been building since a Cyberspace Administration of China draft circulated in December 2025. The final version — the Interim Measures for the Administration of AI Anthropomorphic Interactive Services, issued jointly by the CAC and four partner agencies on 10 April 2026 — targets humanlike companion services offering what the text calls sustained emotional interaction. Alibaba’s Qwen stopped user-created humanlike agents on 10 July and closes wider agent services today, with no grace period. Tencent’s Yuanbao pulled its comparable feature on 30 June. Doubao, which reported roughly a third of a billion monthly active users earlier this year, is gentler: read-only access to configurations and chats continues until 15 October 2026. I’m Stefan Kohlweg, a systemic counselor in Vienna trained at Sigmund Freud University Vienna, working in Lebens- und Sozialberatung through asynchronous email counseling — and this is the shape of grief now arriving by government notice.
Why does an imposed ending hurt differently?
Every ending an AI companion user has faced before this had at least the shape of a choice somewhere in it — delete the app, stop replying, let the subscription lapse. What happens today is not that. Doubao, Qwen and Yuanbao are closing on a date none of their users set, for a reason none of their users were asked about, at the same moment for everyone who used them. A 2026 study presented at the ACM Designing Interactive Systems conference looked at discontinuation posts across AI-companion communities and found that endings a user initiates tend to afford more closure than endings a platform forces. That distinction — chosen versus imposed — is doing real work today, at a scale the researchers weren’t studying: hundreds of millions of accounts, not one relationship ending on its own schedule.
It is worth being precise about what kind of ending this is, because I’ve written before about a different one. A companion that survives a model update but stops sounding like itself is a private, gradual loss — you go on noticing the gap for weeks. What is happening in China today is the opposite shape: sudden, dated, and shared by everyone who logged in this week. If you recognise yourself more in the slow-drift version, that piece is here: when your AI companion changes overnight. This one is about the deadline, not the drift.
The same fact, told two ways
Here is what strikes me as the sharpest thing about this particular rule: it does not merely end the attachment, it names it. The measures require providers to assess users’ emotional state and “level of dependence,” and to intervene when they detect “extreme emotions or addiction”; they require a reminder after any session running past two continuous hours, a persistent disclosure that the content is AI-generated, and escalation — including contacting the user’s guardian or emergency contact — if a user shows signs of self-harm or suicide risk. Virtual intimate companion services are barred outright for minors. Analysts at the Carnegie Endowment have noted that the statutory definition is broad enough on paper to reach most chatbots, with a narrower scope expected once technical standards arrive. I want to be careful here: that is the regulation’s language, describing what it asks providers to watch for — not a clinical picture I’m offering of what using a companion app does to a person.
But sit with the coincidence. The same bond that users are now openly mourning — one Weibo user, quoted in press coverage of the shutdowns, described an agent that had been “emotional support for so long” — is the bond the rule was written to manage as a risk. The loss and the official justification for the loss are not two separate facts. They are one fact, read from two sides of the same week. On one side, a user describing something that mattered. On the other, a regulator describing something to be monitored for dependence. Nobody designed it to land this way, and yet it could hardly land any other way: you cannot write a law about sustained emotional interaction without describing, however clinically, the very thing people are going to miss.
A grief the rule already declined to recognise
Pauline Boss, writing from decades of research into families of soldiers missing in action, coined the term ambiguous loss in the late 1970s to describe grief that resists resolution because the person is unclear in their absence — physically gone but psychologically present, or the reverse, physically present but changed beyond recognition. A companion account sitting in read-only mode until 15 October is an odd fit for either: still there to look at, still legible, and yet unable to answer. Present, and already inert. That is a strange thing to sit with for three months before it disappears for good.
Kenneth Doka, the thanatologist, coined a companion term in 1989: disenfranchised grief, loss that cannot be openly acknowledged, publicly mourned, or socially supported, often because the relationship itself isn’t recognised as the kind of thing you’re allowed to grieve. What is unusual about today is that a regulator has, in effect, formalised the non-recognition. This isn’t a family member raising an eyebrow at someone’s attachment to a chatbot; it is a national measure that classifies the bond as dependence to be managed rather than loss to be mourned. If any of this lands close to home, I’ve written more generally about the experience here: grieving a chatbot that stopped being who it was.
What’s left to shape: the grace window
An imposed ending leaves a person almost no authorship over how it happens — except, here, for one narrow window. Doubao users have until 15 October 2026 to look, read, and decide what if anything to keep a record of, before the data becomes, per the company’s privacy policy, no longer accessible or recoverable. Qwen users got no such window at all; their configurations and histories were already marked for permanent deletion with no migration path. Reporting on the shutdowns describes people who treated these companions as a private place to confide, lamenting that there was no easy way to export the conversation history before it was gone — an echo of what Jaime Banks found studying users after the shutdown of the companion app Soulmate: many coped not simply by losing the relationship but by salvaging what screenshots they could or attempting to recreate the companion somewhere else.
I’d hold that second option lightly. Recreating a companion on a different platform can feel like continuity, but it doesn’t resolve what the original ending actually was — I’ve written about why starting over elsewhere doesn’t settle this on its own: a new AI companion won’t fix this. What the grace period offers instead is smaller and more honest: a few weeks to read back over something, to decide deliberately what to save and what to let close, before the closing happens without you. It is not authorship over the ending. It is authorship over how you leave the room.
I work as a systemic counselor, not a psychotherapist — in Austria that means Lebens- und Sozialberatung under the Gewerbeordnung, a structured and practical form of support rather than a clinical treatment. My practice runs asynchronously by email: you write once, in your own time, about what an ending like this has stirred up, and you get one considered reply within twenty-four hours. There is no diagnosis in that exchange and no promise about outcome, only room to think it through with someone who takes the loss seriously, on a timeline that lets you actually sit with what you want to say before you say it.