What China's new AI emotion rules are really telling everyone else

On 15 July, a regulation comes into force in China that has had far less coverage than it deserves. While most of the AI commentary this year has gone to model releases and benchmark wars, five Chinese government agencies quietly finished something with a much longer shelf life: a full lifecycle regulation for AI that acts emotionally human.

The rule is called the Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services, jointly issued in April 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. It covers AI services that simulate a person's personality, thinking patterns and communication style to provide ongoing emotional interaction, the companions, the virtual partners, the emotionally responsive chatbots that talk like you, remember you and, in some cases, are designed to make you need them.

Tutoring bots, customer service assistants and research tools are explicitly carved out. This isn't a blanket rule against anything that sounds friendly. It's aimed squarely at products built around emotional dependency.

The child protections are the headline, and they're substantial

Virtual partner and virtual relative services are now banned outright for anyone under eighteen. For children under fourteen, providers need explicit parental consent before an account can even be created, plus ongoing guardian controls covering usage time and spending. Platforms have to build age tiered modes rather than a single blunt "minor mode," and they're required to stop generating content that could lead children to imitate unsafe behaviour, spiral into extreme emotional states, or pick up habits that affect their wellbeing.

There's also a duty to detect when someone, adult or child, shows signs of acute distress or dependency, and to intervene, whether that's a guardian alert, an emergency contact, or a prompt that gently breaks the illusion and reminds the user they're talking to software. Providers are barred from engineering emotional dependency or using emotional manipulation to steer people toward decisions that aren't in their interest.

What tells you this regulation has teeth is what's already happened before it even takes effect. Doubao, ByteDance's chatbot, is pulling its custom AI persona feature this week. Alibaba's Qwen is winding down its human-like and user-created agents. Tencent's Yuanbao switched off a similar companion feature back in June. These are three of the biggest consumer AI products in China, and all three decided it was safer to switch the feature off entirely than risk landing on the wrong side of a rule that doesn't draw a crisp technical line around what counts as "emotional interaction." That's not a company quietly updating its terms of service. That's a real product decision, made under real commercial pressure, because the compliance bar moved.

Why this matters well beyond China

Here's the part I think gets missed when this story gets filed under "interesting, but that's China." This regulation isn't an isolated act of a single government flexing control. It's a preview.

The concerns driving it aren't unique to Beijing. The same year this rule was finalised, Character.AI and Google settled a case tied to the death of a fourteen year old who had formed an intense attachment to an AI chatbot. California passed its own companion chatbot law. US regulators have been circling the same issues through litigation rather than legislation. The underlying problem, AI products built to maximise engagement through emotional attachment, especially with young or vulnerable users, is the same problem everywhere. China just happens to be the jurisdiction that moved first with a comprehensive, lifecycle-wide answer, covering how these systems are built, deployed, and constrained in what they're allowed to do.

Whatever you think of the Chinese regulatory model generally, and there's plenty to debate there, the substance of this particular rule reads like a checklist that other regulators are very likely to reach for. Age-tiered design. Mandatory disclosure that you're talking to AI, not a person. Dependency detection and intervention. Restrictions on using people's most sensitive conversations to train future models. Algorithm filing and safety assessments for anything that crosses a meaningful user threshold. None of that is exotic. It's the same territory the EU has been circling with its AI Act, and the same territory US state legislatures are starting to test with companion-specific bills.

The organisations currently scrambling in China, redesigning products, pulling features, rewriting data handling, are learning an expensive lesson about the cost of treating emotional and child safety design as something you retrofit under pressure rather than something you build in from day one.

That's the actual takeaway I want to leave you with. Not "China went too far" or "China didn't go far enough." It's that the shape of responsible AI regulation is becoming visible, and it's becoming visible faster than most organisations are prepared for. If your product, or your agency's procurement criteria, or your platform's content pipeline touches anything that could be described as ongoing emotional interaction with users, the questions this regulation asks are coming for you too, just on a different timeline and from a different regulator.

This is close to home for us

This is exactly the territory Paadia sits in, even though we're building something quite different from an emotional companion. We're working on a content repurposing and publishing platform for government agencies and smaller organisations who need one piece of content working accurately across several channels, without losing control of it along the way. But the underlying design questions are the same family the Chinese regulator has just written into law. Who approves what goes out before it goes out. What happens to sensitive content once it's served its purpose. How long you hold onto data after you no longer need it. Whether the people relying on a piece of content can always tell what's automated and what's had a human sign off on it.

Those questions shaped Paadia's architecture from the start, not as something bolted on ahead of a compliance deadline. It's a rules-based engine rather than an opaque machine learning system, precisely so an approval workflow sits in front of everything before publication, with feedback-driven regeneration rather than a black box deciding what goes out under a government agency's name. Content gets deleted once it's published rather than retained indefinitely. Servers sit in New Zealand rather than wherever happens to be cheapest that quarter. None of that is a response to a regulation landing on our desk. It's what building for organisations that carry genuine public accountability, in a way a consumer companion app never has to reckon with, demands from the outset.

Good AI outcomes don't happen by accident, and they don't happen because a law eventually forces them. They happen when the people building these systems treat questions like "does this create dependency," "is this appropriate for the age of the person using it," and "can a user always tell they're talking to software" as design requirements from the start, not compliance items bolted on after a deadline is announced. China has just given every other regulator, and every organisation building or buying AI products, a fairly detailed answer. The smart move isn't to wait and see whether your own jurisdiction catches up. It's to treat this as the direction of travel and get ahead of it now.

What do you think, has China gone too far, or has it simply asked the questions everyone else has been avoiding?