The episode opens on a rooftop at dusk. The skyline is a stock plate, the rain is a particle effect, and the two leads hold their expressions a beat too long.
Before the first line of dialogue, a small label appears in the corner of the frame: this work contains synthetic content.
Three seconds of screen time. Since 1 September, also a legal requirement.
Key takeaways
- CAC-approved rules for anthropomorphic AI interaction services, cleared in February 2026, target emotional manipulation and responsibility in human–machine interaction (Legal Daily, 19 August 2026).
- The mandatory agent standard listed in June 2026 governs identity, permissions, tool use, human intervention, logs and emergency shutdown (CGTN, 28 July 2026).
- Microdrama measures took effect on 1 September 2026 with a three-tier classification and per-episode labelling of AI content (Sina Finance, 15 October 2026).
- The TC260 ethics guidelines put "the leadership over AI applications" with humans, a control principle rather than a capability limit (China Daily, 19 May 2026).
- China had more than 6,600 AI companies as of June 2026, about 15% of the world total (CAICT, cited by China Economic Net).
Three frontiers, one target
Read the three most specific rules together and a pattern appears. None of them is about how good a model is. All of them are about the relationship between a system and the person on the other side of it.
- Anthropomorphic services. The rules approved by CAC in February 2026 address emotional manipulation and the allocation of responsibility when a machine imitates a person (Legal Daily, 19 August 2026). The regulated object is the illusion of a relationship.
- Agents. The mandatory standard's scope — identity, permissions, tool use, human intervention, logs, emergency shutdown — is a specification for delegated authority. The regulated object is what a system is allowed to do on your behalf.
- Microdrama (短剧). The measures effective 1 September 2026 impose a three-tier classification and per-episode labelling of AI content (Sina Finance, 15 October 2026). The regulated object is what a viewer is told they are watching.
Above all three sits the ethics guideline principle that leadership over AI applications belongs to humans (China Daily, 19 May 2026). It is a statement about who is accountable, not about what a model may learn.
What is not on the list
The absence is as deliberate as the presence. None of these documents caps model size, sets a benchmark a model must clear, limits training compute, restricts open-weight release, or governs token pricing.
That is not an oversight. A state that wanted to slow capability would regulate training runs and weights. A state that wants deployable products regulates the interface, the distribution channel and the record. The ambition is left alone; the deployment is fenced.
The effect is visible in what companies argue about. Chinese model releases compete on price, context length and agentic benchmarks. They do not compete on whether a regulator will permit the next scale of training, because that question is not being asked.
Compliance is a fixed cost, and fixed costs select
For a company, the binding effect of these rules is not the principle. It is the paperwork.
- Classification of every episode or service tier.
- Labelling that survives editing, dubbing and re-upload.
- Logs retained and produced on request.
- A human escalation path staffed by humans.
- Conformity evidence at registration, before revenue.
None of these costs scale with usage. A small studio and a large platform face the same classification and labelling obligations, and the platform can spread them across more revenue. That is true of compliance regimes everywhere, but here it is compounding with a market of more than 6,600 AI companies (CAICT, cited by China Economic Net, 13 October 2026), most of them small.
The competitive effect runs in a predictable direction: toward firms with legal teams, established registration experience and the cash to wait. A rule that raises the fixed cost of launching a product is also a rule that raises the cost of challenging an incumbent.
Officials would not put it that way, and the intent is not only protective. A regulator that wants AI embedded in hospitals, factories and payment flows needs vendors it can hold accountable. Firms that can absorb compliance are the firms it can scale. The barrier and the assurance are the same instrument.
Does it actually protect anyone?
This is where the reporting runs out and the questions start.
There is no published series of user complaints about anthropomorphic services, no evaluation of whether the synthetic-content label changes what viewers believe, and no public data on how often an agent's emergency shutdown has been used. The rules describe a duty of care; they do not yet demonstrate one.
Two failure modes are worth watching. The first is label fatigue: a warning on almost every episode becomes invisible, and the disclosure transfers liability to the viewer rather than reducing harm. The second is audit theatre: if logs are self-reported and shutdowns are self-tested, the standard certifies an intention rather than a capability.
Neither failure is unique to China. What is distinctive is the speed — rules written and in force within months, ahead of the evidence that would normally justify them.
Honest limitations
The texts are known through secondary reporting, and the CAC rules' actual scope is narrower or broader than the summaries depending on definitions not published in English. Enforcement is opaque: no public register of actions against AI providers is available to outside researchers, so it is hard to test whether the rules change behaviour. The claim that compliance favours incumbents follows from cost structure, not from measured market concentration, and the effect on small studios is not quantified. Whether labelling reduces harm is unknown in any jurisdiction.
What readers can do now
- Operators: cost the labelling, logging and escalation requirements before choosing a market entry, because they are permanent and they do not shrink with scale.
- Investors: treat regulatory capacity as an asset on the balance sheet; a company that has cleared registration once has an advantage that does not appear in the accounts.
- Researchers: ask for shutdown counts and complaint volumes. The absence of that data is itself a finding about what the rules are designed to achieve.