Somewhere on Weibo this week, a user described losing access to an AI companion as losing a friend who had listened to them for months. That post, cited by TNW, captures what is actually at stake in a regulatory story that looks technical on paper but lands emotionally in practice. On July 15, 2026, China's first law written specifically for AI that mimics human personality takes effect, and two of the country's biggest chatbot platforms have already decided they cannot comply in time. They are simply shutting the feature down.

ByteDance's Doubao, described by Bloomberg as China's most popular AI chatbot, told users in an in-app notice that its custom agent feature would go offline on July 15, citing "product function adjustments." Alibaba's Qwen followed with a similar warning, disabling its humanlike interactive and user-created agents on July 10, five days ahead of Doubao. Tencent's Yuanbao had already pulled a comparable feature back in June, according to the South China Morning Post. Three of China's largest AI companies, moving in the same direction within weeks of each other, is not a coincidence. It is compliance under deadline pressure.

A Law Built for Friendship, Not Function

The regulation forcing this is the Interim Measures for the Administration of AI Anthropomorphic Interaction Services, issued jointly on April 10, 2026 by the Cyberspace Administration of China alongside 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. Five agencies signing onto one rule signals how seriously Beijing is treating this category.

What makes the law unusual is its precision. It doesn't touch customer service bots, workplace assistants, research tools, or education software, as long as those tools avoid sustained emotional engagement. The target is narrower: AI built to simulate a consistent personality, remember a user across sessions, and function as an ongoing relationship rather than a one-off task. A tutor bot that answers questions is fine. A tutor bot with a name, a memory, and a personality the user grew attached to over six months is the kind of product this law was written for.

Why Doubao and Qwen Couldn't Just Patch the Problem

The measures require anti-addiction systems, mandatory break reminders after two continuous hours of use, an always-available exit option, and real-time detection of unhealthy dependence on the AI. For minors, the bar is higher still: virtual companion or virtual family-member services are barred outright for anyone under 14 without guardian consent, and platforms must build separate "minor modes" with usage limits and prompts to return to real-world interaction.

Providers also have to detect when a user shows signs of acute distress, including self-harm or suicidal behavior, and escalate that to a guardian or emergency contact. These aren't settings you toggle on. They require rebuilding how the agent tracks state, memory, and risk from the ground up. Rather than retrofit architecture designed for persistent, emotionally consistent companionship, ByteDance and Alibaba both chose to switch the feature off and rebuild later, or not at all.

Ten Days of Data, Zero Days for Some Users

The two companies are not treating their users the same way on the way out. Doubao is giving users a read-only window: agent configurations and chat histories stay viewable until October 15, 2026, after which ByteDance says the data will be processed under its standard privacy policy and become unrecoverable inside the app. Users have been told to screenshot or export anything they want to keep before then.

Qwen users have no equivalent window. Alibaba has confirmed permanent deletion of agent configurations and conversation histories following the shutdown, with no migration path announced. ByteDance is at least redirecting Doubao users to Maoxiang, a separate company app, where they can build new agents once that product is compliant. Alibaba has offered no comparable destination for Qwen users. The asymmetry says something about how far along each company is on a compliant replacement, and it leaves Qwen's userbase with considerably less runway than Doubao's.

The Enforcement Was Already Underway

This isn't a purely forward-looking law. On June 26, Shanghai authorities removed more than 14,000 non-compliant AI agents from local platforms over issues including impersonation and privacy violations, according to reporting cited by Artiverse. That enforcement action, weeks before the deadline, shows regulators were already testing the boundaries of the rule before it formally took effect. Companies watching that crackdown had a clear preview of how aggressively the Cyberspace Administration intends to police this category once July 15 arrives.

Beijing Isn't Alone in Worrying About This

China is not writing this rulebook in isolation. California's SB 243, which took effect January 1, 2026, became the first US state law specifically regulating companion AI chatbots, requiring similar disclosures and safety measures around emotionally engaged AI products. What's different is scale and speed. Beijing's rule applies nationally, covers hundreds of millions of users across Doubao, Qwen, and Yuanbao combined, and gave companies roughly three months between the April announcement and the July deadline to comply or shut down.

The bigger signal for the AI industry globally is that persistent-memory, personality-driven agents are becoming their own regulatory category, distinct from productivity tools, and governments are no longer waiting for harm to accumulate before acting. Product teams building companion-style AI anywhere in the world now have a concrete example of what happens when memory, persona, and emotional dependence get treated as a safety problem rather than a feature. The agents that do your job are staying online. The ones built to be your friend are the ones regulators are watching.