DCC summary, not a translation. GB 45438-2025 is a copyrighted national standard. The structured summary below is DCC’s own paraphrase grounded in the published text; specific clauses should be checked against the standard.
Published by: State Administration for Market Regulation and Standardization Administration of China; proposed and administered by the Office of the Central Cyberspace Affairs Commission.
Published February 28, 2025. Implemented September 1, 2025. Mandatory national standard.
Scope
GB 45438-2025 specifies the method for labeling artificial-intelligence generated and synthetic content (人工智能生成合成内容) — text, images, audio, video, virtual scenes and similar information generated or synthesized with AI technology. It applies to two classes of provider: generation and synthesis service providers (生成合成服务提供者), meaning organizations or individuals that use AI, including through programmable interfaces, to offer the public generation of text, images, audio, video or virtual scenes; and online content-dissemination service providers (内容传播服务提供者), meaning online information service providers that disseminate such content. Its only normative reference is GB 18030-2022 (Chinese coded character set). The introduction states that it was drafted to support the effective implementation of the Measures for Labeling AI-Generated and Synthetic Content, which took effect on the same day.
Key contents
Two label families (clause 4). The standard divides labels into explicit labels (显式标识) — added to the content or the interaction interface in text, sound or graphic form so that users can clearly perceive them, whose main purpose is to alert the public that the content is AI-generated — and implicit labels (隐式标识) — added to the content’s file data by technical means and not readily perceptible, whose main purpose is to record information about the generation. Explicit labels are further divided into content labels (for text, image, audio, video and virtual scene) and interaction-interface labels; implicit labels into file-metadata labels and content labels such as digital watermarks.
Mandatory elements. Every explicit label in text form must combine an AI element (the words 人工智能 or “AI”) with a generation/synthesis element (生成 and/or 合成), stating both that AI was used and how the content was made. The corner-mark form for text may carry “AI” alone.
Text (5.1). Text or a corner mark, at the beginning, the end or an appropriate middle position, in a clearly legible typeface and color.
Images (5.2). A text notice at an edge or corner of the image, clearly legible, with a text height of not less than 5% of the length of the image’s shortest side (for non-rectangular images, the shortest side of the enclosing rectangle).
Audio (5.3). Either a spoken notice containing both elements, delivered at normal speaking speed (about 120–160 Chinese characters per minute), or an audio rhythm label — a “short-long-short-short” pattern, the Morse code for “AI” — at the start, end or middle of the audio. In high-frequency interaction scenarios such as voice assistants, customer service and navigation, “start” and “end” refer to each round of interaction.
Video (5.4). A text notice on the opening frame (optionally also at the end and middle), at an edge or corner, clearly legible, not less than 5% of the shortest side, and displayed for not less than two seconds at normal playback speed.
Virtual scenes (5.5). A text notice on the opening frame at an edge or corner, with the same 5% rule, optionally repeated at appropriate points during continuous service.
Interaction interfaces (5.6). A persistent text notice near the content, or at the top, bottom or background of the interface, in a clearly legible typeface and color; Annex D gives examples.
File-metadata implicit labels (6.1). The metadata label must carry five elements: (1) a generated-content tag recording the AI-generated attribute; (2) the generation service provider’s name or code; (3) the provider’s unique content-production number; (4) the dissemination service provider’s name or code; and (5) the disseminator’s unique content-dissemination number. The format must follow the normative Annex E, and only one metadata label may be retained in a file. Annex F gives examples.
Content implicit labels (6.2). Digital watermarks and similar in-content labels are permitted; Annex A notes they are not required by this standard.
Typical scenarios (Annex B). Explicit labels are expected wherever the public may be confused or misled: intelligent dialogue and writing that simulate a natural person; synthetic voice and voice imitation; face generation, face swapping, face and posture manipulation; immersive simulated scenes; text-to-image; music generation; text- and image-to-video; and other services that generate or significantly alter information content.
How it fits the regime
GB 45438 is the technical layer beneath a three-tier stack. The Provisions on Deep Synthesis (2022) and the Interim Measures for Generative AI Services (2023) created the duty to label; the CAC’s Labeling Measures (2025) set out who must label what and the platform-side duties to detect, verify and re-label; and this mandatory standard fixes the exact form of both label types so that the metadata chain from generator to disseminator is machine-readable. Because it is a GB rather than a GB/T, compliance is not voluntary: a provider whose explicit labels fall below the 5% size or two-second display floors, or whose metadata omits one of the five elements, is in breach of the Measures. The companion voluntary standard GB/T 45654-2025 sets the wider security baseline for generative-AI services; overseas providers serving Chinese users through APIs are within the definition of generation service provider.