China's deep-synthesis rule labels the technology, not the character

The persona

China's rule addresses a technology category, 'deep synthesis,' rather than any single account. The Cyberspace Administration of China's Provisions on the Administration of Deep Synthesis of Internet Information Services, promulgated 25 November 2022 and effective from 10 January 2023, defines deep synthesis to include text generation mimicking natural persons, synthetic or altered voice, facial generation or replacement, and immersive simulated scenes, covering any provider offering these capabilities regardless of whether the resulting output is presented as a persona, a filter, or a background effect.

What the documents establish

Article 17 of the Provisions requires providers whose services 'may cause the public to be confused or misled' to display prominent labelling flagging the deep-synthesis content, while Article 16 requires providers to add identifiable markers to generated or edited material without degrading the user experience, and to retain logs. Article 18 separately prohibits deleting, altering or concealing those markers. The State Council's own portal republished the identical text at content_5731431.htm, corroborating both the prominent-labelling requirement and the January 2023 effective date.

Character versus company

A synthesized voice or face presented as a persona's own performance is, under these Provisions, an output the service provider must mark, independent of whatever narrative accompanies the character. The obligation to label, to preserve identifying markers, and to keep logs runs to the deep-synthesis service provider, not to any persona built using that service. A platform hosting a virtual character whose voice was produced by such a service would need to confirm the underlying marker was neither stripped nor concealed, a duty the Provisions place on the technical chain rather than on the persona's presentation.

What to watch

This entry relies on an English rendering of a Chinese-language regulation rather than an official bilingual text, so exact translations of specific terms should be checked against the original before being quoted precisely elsewhere. It is also worth tracking how the labelling requirement interacts with China's later generative-AI rules, since the two regimes were issued separately and their overlap for virtual-persona content specifically is not resolved by either document alone.

• Does the labelling requirement apply to the underlying voice or face model, the finished persona video, or both?
• What log-retention period do the Provisions require, and who can access those logs?
• How does this regulation interact with China's later generative-AI content rules for the same persona-driven output?

The regulation's own structure ties disclosure to a technical capability, not to a character's storyline, which means a persona built on deep-synthesis tools inherits a labelling duty regardless of how its narrative is presented to viewers.