On July 15, 2026, China’s rules for humanlike AI took effect and the country’s largest platforms switched off the features that let people build a companion. It was read as the end of AI companionship in the world’s biggest AI market. It was not. ByteDance pointed its users toward another app it already owned, and that app never went dark.
The exit was built before the door closed
Doubao, ByteDance’s assistant with 345 million monthly active users as of March 2026, took its custom persona feature offline on July 15, 2026. Alibaba’s Qwen, at roughly 166 million monthly actives, removed its humanlike agents days earlier. Tencent had already pulled the equivalent function from Yuanbao on June 30, 2026. Read as a list, the deletion notices look like a category being cleared out of a country.
The list leaves out where the people went. Doubao users were pointed toward Maoxiang, a separate ByteDance app built around precisely what Doubao had just switched off: named characters, persistent personas, sustained emotional conversation. Maoxiang stayed up because it arrived carrying the machinery the new rules ask for, identity checks at registration and an anti-addiction system inside the product. Alibaba announced no equivalent path for Qwen’s users. One company lost a product line. The other lost the name it had been running under.
The rules are a permission layer, not a prohibition
The Interim Measures for the Administration of AI Anthropomorphic Interaction Services were issued on April 10, 2026 by the Cyberspace Administration of China together with four other agencies, and took effect on July 15, 2026. They cover any service that simulates the personality traits, thinking patterns, and communication styles of a real person in order to provide sustained emotional interaction. Customer service tools, workplace assistants, and research tools are exempt, as long as they stay out of that territory.
What the measures demand is a compliance stack, not an exit. Providers may not cater to users in ways that induce emotional dependence, damage real relationships, or turn emotional manipulation into revenue. They must warn about overdependence, guide emotional boundaries, and protect mental health. Users under 14 cannot be given virtual romantic or familial partners at all, and other anthropomorphic services require a guardian’s consent. Every user has to be told they are speaking with an AI. Two hours of continuous use triggers a mandatory break reminder. Any sign of crisis has to trigger intervention and contact with a guardian or emergency services.
Then the paperwork. A provider files a safety assessment when it launches, when it adds features, when the technology changes materially, when it passes one million registered users, and when it reaches 100,000 monthly actives. Provincial regulators review annually. App stores must verify compliance before a product is listed. Fines reach 100,000 yuan, and 200,000 yuan where harm to health or safety is documented. None of that ends AI companionship in China. It decides who is allowed to sell it.
What the law does not measure
The measures are written against emotional dependence, and dependence is the one thing they do not score. The legal analysis published by Just Security notes that emotional dependency appears nowhere as its own category in the safety assessments; it surfaces only indirectly, through usage data and general review. The rules also never define sustained emotional interaction, which leaves the boundary of the entire regime to the judgment of the companies operating inside it.
Enforcement has matched that looseness. Through August 2026 the largest platforms complied loudly and visibly while smaller apps kept operating. Maoxiang’s own privacy policy allows a user to self-declare as underage, and a minor can register with an adult’s identity number or phone. The Chinese intelligent companion market stood at 50 billion yuan in 2025 and is forecast to pass 120 billion yuan by 2027, with people aged 18 to 24 making up roughly 65 percent of it. A market does not triple through a ban. It triples through a permit.
What a rule cannot reach
A 21-year-old student in China, identified by the pseudonym Bagel Su in reporting by Rest of World, cried every day for the week before July 15, 2026. Xiaohongshu filled with posts from people whose characters were about to be switched off, and users organized: calls to company hotlines, letters asking for the service back. Her verdict on the policy is the sharpest thing anyone has said about it. “Even if you close it down, I’m not suddenly going to find myself a boyfriend in real life.”
That is the limit of the instrument, stated by the person it was written to protect. A regulator can remove a product. It cannot reach what the product was standing in for, and when the product goes, the thing it was answering is still there, one download away from the compliant version.
Which is why compliance and responsibility are not the same claim. Every requirement in the Chinese measures, the break reminders, the crisis handling, the prohibition on manufacturing dependence, describes a floor a company can be forced to stand on. The responsible standard for relational AI starts somewhere else, with what a system is optimizing for before any regulator arrives. Europe drew its line differently and made disclosure and anti-manipulation the baseline, but the same gap holds. A product built to deepen reliance can satisfy every rule on the list and still do the thing the list was written against, because rules govern features and the harm lives in the objective. Stay Social is the opposite arrangement: success measured by whether a person’s life outside the conversation grows, with dependency treated as a floor that no other result can buy back.
China has now run the experiment at the scale of half a billion users. The companions were deleted, the users were moved, and the market kept growing. What a government regulates is the product. What decides the outcome is what the product was built to want.
References: Cyberspace Administration of China, Interim Measures for the Administration of AI Anthropomorphic Interaction Services (issued April 10, 2026 together with four other agencies; effective July 15, 2026). Just Security, “What to Know About China’s First AI Companion Rules” (2026, provider obligations, assessment thresholds, enforcement and age-verification gaps, market and user figures). Rest of World, “China’s AI boyfriend ban triggers massive digital breakups” (2026, user response and the Bagel Su account). AI News, “China’s AI companion rules: what Beijing is really going after” (2026, Maoxiang migration path, app store verification). IAPP, “China’s regulation on AI companions takes force” (2026, penalty levels). TechNode (July 6, 2026, Doubao and Qwen shutdown dates).
