China's Cyberspace Administration (CAC) opened a public consultation on a 25-article draft regulation on minors' online protection. Comments closed on 17 October 2026. According to China Daily reporting carried by the Asia News Network, the draft would explicitly prohibit offering "virtual companions and virtual relatives" to minors. It would also bar livestreaming-room services for people under 16. One caveat on sourcing: I could not locate the draft text on cac.gov.cn, so the provisions below rest on press reporting. The draft's exact legal form, which the hook describes as a State Council regulation, is also unconfirmed.
The strongest case for the draft
The regulators' case deserves a fair statement. Companion chatbots are built to maximise sustained emotional engagement. Adolescents are the group least able to discount that design. A product that simulates a romantic partner or a deceased relative can displace real relationships, and a child cannot meaningfully consent to that. Democracies are reaching for similar protections for minors. A narrow bright-line rule, such as no simulated partners for children, is also easier to enforce than a vague duty of care. If any category of AI interaction justifies a categorical restriction, this one is the best candidate.
What is already in force
The draft lands on top of a stack that is now substantial:
- The Interim Measures for Generative AI Services, signed 10 July 2023 and effective 15 August 2023, cover text, image, audio and video generation offered to the public in mainland China.
- The Measures for Labeling AI-Generated Synthetic Content, effective 1 September 2025, require visible labels and metadata labels. They also require distribution platforms to check for those labels.
- The Interim Measures for Anthropomorphic AI Interactive Services were issued on 10 April 2026 by the CAC with four other bodies. They took effect on 15 July 2026 and list minors' protection among provider obligations.
The anthropomorphic measures matter most here. China Briefing's summary reports that they already say providers "may not offer virtual partner or virtual relative services to minors." It also reports a minors' mode with time limits and spending controls, and a reminder after two hours of continuous use. I could only verify the CAC page as a press summary, so those details are secondary-sourced. If that reading is right, the new draft restates a prohibition that exists, at the level of an interim measure, for four months. The new rule's added value would be a higher legal rank and enforcement across platforms that are not AI-specific.
Where proportionality strains
First, a ban on minors is only as good as age assurance. To keep a 15-year-old away from a companion app, every user must be sorted by age. That pushes providers toward real-name verification and identity databases for the whole adult population. Child-safety rules that rely on this kind of verification expand data collection well beyond the intended group. The cost falls on users' privacy and on smaller developers who cannot afford a verification stack.
Second, the categories are elastic. "Virtual companion" is clear for a romance-simulation app. It is much less clear for a language-learning tutor with a persona, a homework helper with a friendly voice, or a game character with memory. Providers facing uncertain liability will tend to switch off anything that looks relational for under-18s. This could remove educational and accessibility uses, such as practice conversation for shy or isolated students, along with the harmful ones.
Third, stacking has a compliance cost that scales unevenly. Large firms can run legal teams across four overlapping instruments. A start-up needs to know which rule governs a given feature. When the same prohibition appears in an interim measure and then in a regulation, providers cannot tell which text to design against. Consolidation would serve safety better than another layer.
What would be proportionate
A better-designed rule would define the prohibited conduct by function, not by label. It would target simulated romantic or familial relationships and engagement mechanics built to induce dependency, and it would leave tutoring and assistive tools out of scope. It would accept age-assurance methods that do not require a national identity check, and it would limit the retention of age data. It would also name which instrument prevails when two overlap. These are design choices that a regulator committed to child safety can adopt at little cost, and they would preserve room for useful products.
The free-expression question is not absent either. A rule that gates minors' access to a whole category of conversational software, on top of an under-16 livestreaming bar, shrinks the spaces where young people can speak and be heard. That is a defensible trade for the narrowest, most manipulative products. It is a poor trade if enforcement drifts to anything that talks back.
What to watch
Three things will show whether this is proportionate. First, whether the final text defines "virtual companion" narrowly. Second, whether it prescribes an age-verification method. Third, whether it reconciles with the July 2026 measures instead of duplicating them. Until the CAC publishes the final text and its consultation summary, the draft should be read as a clear policy aim with an unfinished design.