China's deep synthesis rules take effect

Background

"Deep synthesis" (深度合成) is the term the provisions use for content produced with generative algorithms. The rules define deep synthesis technology as technology that uses deep learning, virtual reality and similar generative algorithms to produce text, images, audio, video and virtual scenes^[2]. That covers a broad family of outputs — machine-written text, synthetic speech, generated faces, edited video and computer-built scenes — rather than a single kind of product.

The provisions were adopted on 3 November 2022 and published as Order No. 12, dated 25 November 2022^[2]. Three departments released them jointly: the Cyberspace Administration of China, the Ministry of Industry and Information Technology and the Ministry of Public Security^[1]. They took effect on 10 January 2023^[1][2].

What the provisions require

The core obligation is labelling. A provider must add labels that do not affect use to information generated or edited with its service^[1]. The phrase "do not affect use" matters: the mark is meant to inform, not to disable or degrade the output, so a synthetic image or a generated voice can still be watched, heard and shared.

A second, stricter tier applies to particular service types. For services offering intelligent dialogue, synthetic voice, face generation or immersive realistic scenes, providers must add prominent labels, so as to avoid public confusion or misrecognition^[1][3]. These are the formats in which output comes closest to standing in for a person: a voice that sounds like someone, a face that looks like someone, a conversation that reads as if a person were holding it. The rules therefore ask for a mark that is hard to miss rather than one tucked away out of sight.

The provisions also protect the labels themselves. No organisation or individual may use technical means to delete, tamper with or conceal them^[1]. Without such a clause a labelling duty would be easy to work around: whoever could strip the mark could pass synthetic material off as real, and the obligation placed on the provider would do little for the audience it is meant to protect.

Obligations beyond labels

Three further duties sit with providers. They must verify the real identity of users^[1], which ties an account to a person and makes it possible to establish who produced or distributed a piece of content. They must establish mechanisms for refuting rumours and for complaints and reports^[1], giving users and third parties a channel through which synthetic material can be challenged and corrected. And providers and technical supporters that have public opinion attributes or social mobilisation capacity must complete filing and update or cancel it as required^[1] — in effect a register of services whose output can reach and move a large audience.

Why it matters

Taken together, the provisions move the burden of disclosure onto the side that makes the content. A viewer, listener or reader no longer has to work out from scratch whether a video, a voice or a passage of text came from a machine; the service that produced it is required to say so, and to say it plainly where the risk of confusion is highest^[1][3]. Identity verification and complaint mechanisms add a route for tracing and correcting material after it has spread^[1], and filing gives supervisors a list of the services that matter most in shaping public discussion^[1].

The approach is regulatory rather than prohibitive. The provisions do not bar synthetic content; they attach conditions to the services that generate it — labels, real-identity checks, rumour-refutation and complaint channels, and filing — so that deep synthesis can be used while its origin remains visible. Because the duties are written at the level of the provider, the rules place the work of disclosure on companies running the services rather than on the public trying to spot a fake.