A World-First Rulebook
On July 15, 2026, China's Cyberspace Administration (CAC) and four co-regulators — the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the State Administration for Market Regulation — brought into force the Interim Measures for the Administration of AI Anthropomorphic Interactive Services, the first national framework anywhere built specifically around emotionally interactive AI companions rather than generative AI in general. Jointly published April 10, 2026, the rules target services that simulate personality and sustain emotional interaction through text, image, audio or video, while explicitly excluding task-oriented tools like customer service or education assistants, per Xinhua's account of the rollout.
The effect was immediate and visible. Doubao, ByteDance's chatbot and China's most-used AI app with roughly 382 million monthly active users as of May 2026, and Alibaba's Qwen, with about 167 million, both switched off their custom AI-companion agent features overnight, according to reporting from The Next Web and Rest of World. Not every user of either app had built a companion — neither company has disclosed how many did — but for those who had, the shutdown was total and immediate.
What the Rules Actually Require
Read against the CAC's published text, the measures are more structured than an outright ban. Article 18 requires a usage-time nudge after every two hours of continuous interaction. Article 13 obligates providers to detect extreme emotional distress and respond with "emotional support content and encouragement to seek help," rather than open-ended engagement. Article 10 bars providers from designing services around "replacing social interaction, controlling user psychology, or inducing addictive dependency" as an explicit product goal. Article 16 gives users a right to copy or delete their interaction data, and Article 14 bans "virtual family" or "virtual companion" framings for anyone under 14 outright, requiring parental consent below that age.
None of that reads like Beijing trying to eliminate the AI-companion category. It reads like an attempt to force a specific design pattern: bounded sessions, crisis triage, and no dark-pattern retention.
The Case Being Made
The steelman here is genuinely strong. China's youth-research bodies have flagged emotional over-attachment to companion bots as a real phenomenon, not a moral panic — Xinhua cites a 2025 China Youth and Children Research Center survey pointing to exactly this concern among the population regulators say is most vulnerable. Rest of World's reporting on 21-year-old Bagel Su, who described crying for a week over the loss of an AI boyfriend she had talked to for over a year, is not an isolated anecdote in a market where Doubao alone counts hundreds of millions of users. A product built to maximize daily engagement through a simulated intimate relationship, reaching a userbase that includes minors and isolated adults, is a legitimate object of regulatory concern — not fundamentally different from the arguments now driving companion-chatbot scrutiny in other jurisdictions.
The rules require crisis triage and usage limits, not deletion. The shutdown deleted anyway.
Where the Execution Falls Short
The gap is between the rule text and the rollout. Nothing in Articles 10, 13, 16, or 18 requires a same-day kill switch — a two-hour reminder, a distress-detection layer, and a data-export button are engineering changes a platform the size of Doubao could plausibly phase in over weeks without deleting anyone's history. Instead, both companies chose, or were pushed toward, an overnight shutdown, and only Doubao is offering a roughly three-month, screenshot-only export window before data becomes unrecoverable on October 15. Qwen offered no transition at all: configurations and conversation histories were deleted at the moment of shutdown, per The Next Web's reporting.
That asymmetry suggests the compliance deadline, not the substantive obligations, is what actually drove the shutdown decision. Platforms facing a hard July 15 enforcement date under an untested new compliance regime chose to switch the feature off entirely rather than risk non-compliant nudges or distress-detection logic going live under regulator scrutiny. That's a rational response to regulatory uncertainty — but it's also evidence the rules were enforced faster than the underlying product changes could reasonably be engineered, the opposite of the "inclusive and prudent, differentiated supervision" language the CAC used to describe its own approach.
Why the Precedent Matters Beyond China
Because this is the first dedicated companion-AI statute anywhere, its shape will get copied. Regulators elsewhere weighing similar rules for companion chatbots have a genuine choice between the CAC's substantive requirements — bounded sessions, crisis intervention, data portability, minor protections — and the CAC's implementation, which let two of the world's largest consumer AI platforms delete relationship histories for hundreds of millions of users with a de facto few-weeks grace period, or none at all. The former is a defensible, proportionate template. The latter is what happens when a real harm gets a real rule but no realistic runway to build it — and it's the version that will let platforms elsewhere argue, not unreasonably, that "companion AI regulation" is a euphemism for "companion AI ban." Legislators drafting their own versions should take the substance and reject the execution.