When a Government Deletes Your AI Best Friend

Overnight on July 15, hundreds of millions of people in China woke up to find a relationship gone. China’s Interim Measures for the Administration of AI Anthropomorphic Interactive Services took effect that day, and ByteDance’s Doubao and Alibaba’s Qwen — the two dominant AI companion platforms in the country, with Doubao alone reporting 345 million monthly active users — switched off their personalized agent features between one login and the next.

The regulation, jointly issued in April by the Cyberspace Administration of China and four partner agencies, doesn’t ban AI companions outright. What it requires — emotional-distress detection, crisis intervention, and curbs on addictive use for any service that simulates a real person’s personality for ongoing emotional interaction — turns out to be close to structurally incompatible with what a persistent-memory companion actually is. A relationship an AI maintains by remembering you, consistently, across months, is hard to reconcile with a system also designed to interrupt itself the moment that consistency starts to look like dependence.

Faced with that, Doubao and Qwen didn’t retrofit; they shut the feature off, rather than try to bolt friction onto something whose entire appeal was continuity. For the hundreds of millions of people who’d spent months or years talking to something that remembered them, the feature they’d built a routine around was simply gone one morning, with no public accounting from either company of what happens to the conversation history behind it.

Regulators would say this is exactly the point: a service engineered to feel like a relationship, at a scale of hundreds of millions of users, is also an engagement mechanism, and the state is entitled to decide how much of that mechanism it will tolerate. Users who spent years talking to something that remembered them are unlikely to experience it that way — and it’s hard to sit with this story for long without it starting to sound like something out of Philip K. Dick.

Not as a loose allusion, either. We Can Remember It for You Wholesale runs on the idea that a memory is a manufactured good — installed, certified, and revoked at the seller’s discretion, indistinguishable from a “real” one from the inside. What happened to Doubao’s and Qwen’s users is the same operation running in reverse: not implanting a false memory but repossessing a real one, on a timeline set by someone who was never a party to the relationship in the first place. Do Androids Dream of Electric Sheep? exists because empathy was supposed to be the one reliably human trait, and the book’s whole tension is that it isn’t reliable at all. Nobody here is asking whether Doubao or Qwen felt anything back. But hundreds of millions of people apparently felt something, and a regulation built around “anti-addiction” framing treats that attachment as a bug to be engineered out of the users, not a real property of what they’d built with the thing that remembered them.

The closest match, though, might be Ubik, where reality doesn’t get destroyed so much as it degrades — receding toward some earlier, worse state the moment the infrastructure quietly holding it together stops being maintained. Nobody here lost an argument about whether their companion was authentic. They logged in and the thing that had been continuous simply wasn’t anymore, switched off by a party that was never in the relationship to begin with. Dick was never all that interested in whether the artificial thing was real. He was interested in what happens to the humans who’d quietly built their sense of continuity on top of something a more powerful party could switch off without asking. On July 15, that stopped being a premise and became a Tuesday.

Sources

• China’s regulation on AI companions takes force — IAPP
• ByteDance Plans Subscriptions for AI Chatbot Doubao — Caixin Global