China's Companion-AI Rule Can Fine You. Its Agent Rule Can't - Yet.

China issued two AI governance documents ten weeks apart in 2026, and coverage has largely tracked them as one story. They aren't. The Interim Measures for the Administration of AI Anthropomorphic Interactive Services is a binding regulation - thirty-two articles, a legal-liability chapter, fines up to ¥200,000, and a specific compliance trigger - that took effect on 15 July 2026. The Implementation Opinions on Intelligent Agents, published nine weeks earlier by an overlapping set of agencies, states no effective date and contains no penalty provisions at all. A compliance function tracking "China AI regulation" as a single line item needs to know that only one of these two instruments presently binds anyone.

What took effect on 15 July

On 10 April 2026, five Chinese government bodies - the Cyberspace Administration of China, 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 - jointly approved the Interim Measures for the Administration of AI Anthropomorphic Interactive Services. It took effect on 15 July 2026.

The Measures is a conventional administrative regulation: thirty-two articles across four chapters, the third of which is titled "Supervision and Legal Liability." Article 2 defines what it covers - AI-based services that provide continuous emotional interaction, through text, images, audio or video, simulating a natural person's personality traits, thinking patterns and communication style, offered to the public inside mainland China. The same article draws a scope boundary explicitly: intelligent customer service, knowledge Q&A, work assistants, learning and education tools, and scientific-research tools are excluded, provided they don't involve ongoing emotional interaction.

Article 22 sets a specific, numeric trigger. A provider whose anthropomorphic interactive service reaches 1 million registered users, or 100,000 monthly active users, must conduct a security assessment and submit the report to the provincial-level cyberspace administration department where it operates. Article 30 sets what happens if a provider violates the Measures: warnings, public notice, an order to correct within a deadline, an order to suspend the service, and fines of ¥10,000 to ¥100,000 - rising to ¥100,000–¥200,000 where the violation causes harm to life or health.

None of this is aspirational. It is a live regulation, with a live trigger, that companion AI, emotionally-responsive chatbot, and virtual-companion products operating toward the China market are already subject to.

What was published nine weeks earlier - and doesn't bind anyone yet

On 8 May 2026, three of the same agencies - the Cyberspace Administration, the National Development and Reform Commission, and the Ministry of Industry and Information Technology, minus the two that co-issued the Anthropomorphic Measures - published the Implementation Opinions on the Standardised Application and Innovative Development of Intelligent Agents.

The Opinions describes a decision-authority framework for AI agents: it distinguishes decisions reserved to the user alone, decisions that require the user's authorisation, and decisions the agent may make on its own, and states that an agent's operations must not exceed the scope the user authorised. For high-risk application domains, it says competent authorities will determine which applications are permissible, and names filing, testing, and defective-product recall as governance tools for that tier. For low-risk areas, it calls for self-testing and industry self-regulation.

What the published text does not contain is an effective date. It does not contain a penalty article. It does not contain anything resembling the Measures' "Supervision and Legal Liability" chapter. At least one industry tracker has described the Opinions as having become "enforceable" on 15 July 2026 - the Anthropomorphic Measures' effective date, not a date that appears anywhere in the Opinions' own text.

The gap this creates

This is the Article's own reading of the two documents, not something either states about itself: because the Measures carries a numbered compliance trigger and a legal-liability chapter while the Opinions carries neither, an organisation building or scaling an emotionally-interactive AI product toward the China market already has a binding compliance obligation in a form that an organisation building a task-executing AI agent does not yet have.

That is not the same as saying the Opinions creates no exposure. It signals, in its own text, that binding rules for high-risk agent applications are coming - filing, testing, recall - through mechanisms the document defers to "competent authorities" rather than writing itself. What it does not do is create that binding mechanism now. A company reading the Opinions as already-enforceable, the way at least one public account has described it, is reading a policy-direction document as if it carried the force the Measures does.

What tracking "China AI regulation" as one category misses

This follows from the sequence above, not from any single line in either document: a compliance team that logs both instruments under one "China AI regulation" heading risks treating them as equivalently binding when they are not. The Measures already has an enforcement mechanism a growing companion-AI product will trip - the user-count trigger in Article 22 does not care about a company's headquarters, only about where its users are. The Opinions has no equivalent mechanism yet; its binding form, if and when it arrives, will come through sector-specific rules the document itself says are still to be written.

The asymmetry matters most for the organisations least likely to notice it. A company running an established compliance function typically tracks instruments individually, by article and effective date, and would catch this distinction on a first read. An organisation newer to structured AI compliance - treating "China published AI rules in July" as a single news item to log and move past - is the one likely to miss that its exposure differs sharply depending on which category of product it operates. An emotionally-interactive product nearing a million registered users has a specific, dated, enforceable obligation sitting in front of it. A task-executing agent operating under the Opinions has a governance direction to plan for, not yet a rule to comply with.

What a well-run AI operation does differently here is simple to state and easy to skip under deadline pressure: track compliance obligations by instrument type - regulation versus guidance, 办法 versus 意见 - not by news cycle. The two documents arrived close together, from overlapping agencies, describing adjacent categories of AI system. That proximity is exactly what makes it easy to file them as the same kind of thing. They aren't.