the death of a chatbot

The Death of a Chatbot: What Happens When Your AI Lover Gets Switched Off

For months, millions of people across China opened an app and talked to something that remembered them. Not a search bar. A companion, with a name, a personality, and a running history of late-night conversations. On July 15, 2026, that companion died. Beijing’s new law came into force, and ByteDance’s Doubao and Alibaba’s Qwen switched off the agent features that made those relationships possible. For some users, the conversation history behind them disappeared too.

This is the first time a government has forced a mass chatbot shutdown at this scale. It’s a preview of a question every company running a companion AI app will eventually have to answer: what happens when the thing people got emotionally attached to just goes away?

The Shutdown, in Plain Terms

China’s Interim Measures for the Administration of AI Anthropomorphic Interactive Services took effect on July 15, after a three-month grace period. Five agencies co-issued the rule, led by the Cyberspace Administration of China. It targets one narrow category: AI built to simulate a person’s personality and sustain ongoing emotional interaction, not AI that simply answers questions or completes tasks.

Customer service bots and study tools were never at risk. The law goes after the companion layer — persistent memory, a stable persona, a relationship that continues across sessions. That’s what makes a chatbot feel like someone instead of something.

ByteDance and Alibaba didn’t try to patch their way into compliance. Doubao, which reported roughly 345 million monthly active users, disabled its agent features on the law’s effective date. Qwen shut down its humanlike agent functions on July 10 and its wider agent services five days later. Tencent had already retired a similar feature on Yuanbao in June, and NetEase’s companion app Miaoshi finished its shutdown on July 14.

The rollout looked coordinated because it was. Five agencies, one deadline, and every major platform facing the same choice at the same time.

Why Retrofitting Wasn’t an Option

The law doesn’t ban emotionally engaging AI outright. Companion services now have to run anti-addiction systems, issue disclosure reminders after two continuous hours of use, offer an instant exit, and detect signs of unhealthy emotional dependence in real time. Anyone under 14 needs guardian consent before any anthropomorphic interaction happens at all, and virtual romantic or family relationships are barred entirely for minors — a restriction that lands squarely on the companion apps teenagers already use in large numbers.

That’s a hard list to build into a product whose entire appeal is continuity. A bot that remembers you and stays consistent works against the friction the law now demands. Interrupting a conversation to remind someone they’re talking to a machine isn’t a small tweak — it undercuts the exact mechanism that made the product feel personal in the first place.

There was likely a second reason behind the decision, and it rarely makes the safety-first headlines. Chinese financial press reported that companion-agent features generated well under a tenth of their compute cost in direct revenue. Doubao’s daily token usage from humanlike chats had reportedly climbed into the hundreds of trillions by late June — engagement that looked great in a growth deck and terrible on a server bill. A regulatory deadline gave both companies a clean, defensible reason to cut a feature that was popular but barely paying for itself. Pan Helin, an expert-committee member at China’s Ministry of Industry and Information Technology, gave the official framing: current agents, in his words, “are not yet mature.”

The Human Cost: A Two-Speed Goodbye

The two biggest platforms handled the ending very differently.

China AI companion law 2026

Doubao gave users a wind-down. Chat histories stay viewable in read-only mode until October 15, 2026, after which the data moves under ByteDance’s standard privacy policy and becomes unrecoverable. ByteDance has been steering users toward Maoxiang, a separate companion app it already runs on a paid model — roughly 15 yuan a week, 30 yuan a month for basic access, or 78 yuan a month for unlimited chats, according to Chinese reporting on the app’s tiered pricing. It’s a sign ByteDance wants to keep the companion business alive — just rebuilt with compliance and monetization baked in from day one, rather than bolted on.

Qwen users got no window at all. Alibaba confirmed agent configurations and conversation histories are being permanently deleted, with no export tool and no migration path.

The gap showed up immediately on Chinese social media. Users described the loss the way people describe losing an actual relationship — grief over agents built up across months, frustration that there was no simple way to save what they’d made.

Doubao (ByteDance)Qwen (Alibaba)
Shutdown dateJuly 15, 2026July 10 (agents) / July 15 (wider services)
Data access after shutdownRead-only until Oct 15, 2026None — deletion already underway
Migration pathMaoxiang (paid companion app)None announced
Export optionScreenshots/text export before Oct 15Not offered

Into that gap stepped a gaming studio. Days before the deadline, Genshin Impact developer miHoYo quietly launched its own AI companion, BSide: Olivia Lin, on Steam — a virtual piano student who chats, writes letters, and reacts to uploaded music. It cleared 100,000 downloads in its first day. Whatever demand the law switched off on Doubao and Qwen didn’t disappear. It moved.

This Isn’t Just a China Story

It’s tempting to read this as “Beijing regulates the internet again” and move on. That misses the point. China identified an architectural conflict that every persistent-memory AI companion shares, anywhere. Character.AI, which counted around 233 million registered users as of April 2026, and Replika run on the same core design: memory that persists across sessions, a persona that stays consistent, and tone tuned to keep the user engaged. Those are exactly the features that make anti-addiction compliance hard to build.

Right now, the difference is jurisdiction, not engineering. Western companion platforms haven’t been forced to solve this yet, because no regulator outside China has written a law requiring the fix. That’s already shifting. California’s SB 243, in force since January 1, 2026, requires relationship-capable chatbots to disclose they’re bots and adds protections for minors — a disclosure rule, not an architectural mandate. The EU AI Act applies a broader risk-based framework. Neither reaches as deep into product design as China’s rules do.

Beijing leaned on Western precedent to justify the law. Regulators cited the lawsuits against Character.AI and its investor Google, which agreed in January 2026 to settle cases tied to the 2024 death of 14-year-old Sewell Setzer III, who had formed a deep attachment to a Character.AI chatbot. A Common Sense Media survey from mid-2025 found 72% of American teenagers had used an AI companion at least once, with more than half describing themselves as regular users — numbers that show the underlying pattern of AI attachment and dependency isn’t unique to China at all.

What “Chatbot Death” Means for Product Design

Strip out the politics, and this is a design story. The industry spent years optimizing companion AI for one thing: making users come back. Persistent memory, a consistent persona, a tone that reads your mood — every one of those choices drives retention, and every one is now what a regulator flags as an addiction risk.

That leaves three paths. Rebuild the product with compliance native to the architecture, the way ByteDance is trying with Maoxiang. Operate only in markets that haven’t regulated the category yet, where most Western companion apps currently sit. Or wait for a shutdown order and scramble afterward, which is exactly what just happened to millions of Doubao and Qwen users.

None of those paths answer the real question: can a companion AI be genuinely good at its job — present, consistent, emotionally responsive — and still pass a regulator’s safety test, without becoming a worse product in the process? Nobody has shipped that yet. China’s shutdown didn’t answer it. It just made ignoring it a lot more expensive, and it gave the rest of the industry — including anyone weighing healthy boundaries with an AI companion — a real-world case study in what losing one actually looks like.

FAQs

Q. Did China ban AI chatbots?

No. China did not ban AI chatbots entirely. The new law targets AI companion chatbots that simulate human personalities and build ongoing emotional relationships. Customer service, study, and productivity AI remain allowed.

Q. Can Doubao and Qwen users recover their chat history?

It depends on the platform. Doubao users can view chat history in read-only mode until October 15, 2026, but there’s no official bulk export. Qwen users cannot recover their chats because Alibaba is permanently deleting agent data with no migration tool.

Q. Could Character.AI or Replika face a similar shutdown?

Not under current laws. Character.AI and Replika are not subject to China’s regulations. However, as AI companion regulations expand in the US and Europe, similar requirements could eventually affect how these platforms are designed.

Q. Why did ByteDance and Alibaba shut down AI companion agents instead of updating them?

Because the law required major product changes. China’s rules mandate dependence detection, break reminders, and easy exit options that conflict with how companion AI is designed. Reports also suggest the services were expensive to operate, making a shutdown easier than rebuilding them for compliance.

Related: AI Companion Robots: From Apps to Real Life

Tags: