On April 3, 2026, the Cyberspace Administration of China (CAC) opened public comment on the Administrative Measures for Digital Virtual Human Information Services (draft), with feedback due by May 6. The measures extend China's already dense generative-AI rulebook to a new object: the synthetic human persona — virtual news anchors, AI pop stars, motion-capture livestreaming avatars, and AI-driven customer-service faces. The draft requires every such persona to carry a continuous, conspicuous "digital human" label, demands consent before cloning a real person's face or voice, bars digital humans from defeating identity checks, and restricts virtual-companion services aimed at children.
The document deserves to be read on its own terms. It is not the December 2025 draft on "human-like interactive AI," which targets chatbots and emotional dependency; this one treats the digital human as an information medium and borrows the logic China already applies to livestreaming and online content platforms.
The case for acting
Start with the strongest version of the regulator's argument. Face-swapping and voice-cloning have moved from novelty to fraud vector. A synthetic anchor that looks and sounds like a trusted broadcaster, or an avatar built from a real person's likeness without consent, is a genuine harm — to the impersonated individual, to viewers who cannot tell operator from algorithm, and to the integrity of identity-verification systems that increasingly gate banking and government services. A disclosure-and-consent regime answers a real coordination problem: no single creator will label voluntarily if competitors do not, so a baseline rule can be welfare-improving rather than merely restrictive.
Seen this way, parts of the draft are not just defensible — they are the kind of narrow, harm-targeted rule we should want. Article 8 prohibits offering a digital human identifiable as a specific real person, including a "highly similar" likeness or voice, without consent. Article 11(5) bars using digital humans to evade facial recognition, voice recognition, or other authentication where law requires real identity. Both attach liability to concrete, demonstrable wrongs: non-consensual impersonation and verification fraud. That is proportionate regulation.
Where transparency becomes control
The trouble is what the draft bolts onto that sound core. Article 11(1) prohibits content that "endangers the nation's security, honor, and interests," incites subversion of the socialist system, or — per the CAC notice — "distorts historical facts" or "violates socialist core values." Article 22 requires providers with "public-opinion attributes or the capacity to mobilize society" to undergo state security assessments. These are not anti-fraud provisions. They are speech-governance provisions, and they import the discretionary, politically defined categories that run through China's 2021 Algorithmic Recommendation Provisions and 2023 Deep Synthesis rules. A virtual persona becomes a regulated broadcast channel, with the regulator holding a veto over what it may say.
This is the recurring pattern in Chinese AI rulemaking: a legitimate transparency tool fused to a content-control mandate. The 2023 Interim Measures for the Management of Generative AI Services (effective August 15, 2023) did it; the Measures for Labeling of AI-Generated Synthetic Content (effective September 1, 2025) did it. The labeling regime — explicit on-screen marks plus implicit metadata tags — is, in isolation, a reasonable provenance standard not far from what the EU AI Act's Article 50 transparency duties contemplate. But it travels inside a system whose ultimate purpose is to keep synthetic media inside state-defined bounds.
The cost of the always-on label
Even the labeling rule, taken literally, raises proportionality questions. Article 13 requires a conspicuous "digital human" notice displayed continuously throughout the presentation. A one-time, clear disclosure serves the public's interest in knowing what they are watching. A permanent on-screen overlay is a heavier hand — it degrades the product, treats every use case (a virtual sign-language interpreter, a museum guide, a corporate explainer) as equally suspect, and offers little marginal benefit once the viewer has been told. Good disclosure law calibrates to the risk of deception; a blanket persistent watermark does not.
The compliance architecture compounds this. Logging duties, security-monitoring and emergency-response systems, anti-addiction mechanisms, and assessment filings are manageable for a state broadcaster or a platform giant. For an independent creator running a single virtual influencer, they are a moat that favors incumbents — the opposite of what an emerging creative medium needs.
What's genuinely worth protecting
The minor-protection provisions are the draft's most sympathetic. Article 10 bars offering children "virtual relatives" or "virtual companions" and services that induce excessive spending or unsafe imitation; the CAC notice requires guardian approval to model the likeness of anyone under 14. The risk of monetized parasocial dependency aimed at kids is real, and a targeted limit here is defensible. The concern is scope creep: "induce religious belief" and "violate socialist core values" sit in the same provision cluster, blurring child safety into ideology.
The exportable half
The honest verdict is that this draft is two regulations wearing one title. One is a competent synthetic-media transparency-and-consent framework — label it, get consent to clone a real person, don't use avatars to beat identity checks, protect minors from manipulative companion products. Those principles are exportable; democracies wrestling with deepfakes could adopt the consent and anti-fraud pieces tomorrow. The other is a content-control and public-opinion-management layer that other jurisdictions should decline to copy, because its operative terms are defined by the state and enforced at its discretion.
Foreign and domestic operators will have to comply with the whole package. Policymakers elsewhere should take the separable good — narrow disclosure tied to demonstrable harm — and leave the rest. The measure of a virtual-human rule is whether it stops fraud and impersonation without appointing the regulator as editor. This draft does the first and the second at once, and the second is the part that travels worst.