Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act — C(2026) 5054 final, Annex

(60) Article 50(2) AI Act provides an exhaustive list of the modalities of the synthetic content covered by the transparency obligation: audio, image, video or text. This also includes multimodal content, that is, content made up of a mix of these modalities. For the purpose of Article 50(2) AI Act, the different modalities are to be interpreted in line with practical and technological developments that may evolve over time. For the time being, they should be understood as follows:

(26) As defined in Article 3(1) AI Act.

- Text refers to discrete symbolic content composed of characters, or numbers arranged in a particular sequence, that are capable of being read and interpreted semantically by humans.

- Images refer to static spatial representations encoding visual information that are capable of being seen by humans.

- Audio refers to a time-varying signal encoding sound that is capable of being perceived through hearing by humans. This may cover speech, instrumental music or other audio signals.

- Video refers to a time-based sequence of images, which may be synchronised with audio, capable of being seen by humans. In that case, both video and audio content must be marked and detectability must be ensured, as set out below.

(61) AI systems generating or manipulating 3-D images, 3-D videos and 3-D audio, as well as virtual, augmented or mixed reality, fall within the scope of Article 50(2) AI Act. ‘Virtual reality’ covers AI technologies that enable users to experience and interact with a computer-generated environment simulating physical presence and that allow real-time user interaction with that environment. As virtual and augmented reality is made up of a time-based sequence of images, optionally synchronised with audio, that represents visual change over time, it can be regarded as ‘video’ content. Relevant ‘output’ to be marked is the final form of the individual AI-generated asset with special exceptions for the real-time ephemeral interactive sessions under point 88 below. Digital artefacts used to generate 3-D images, 3-D videos and 3-D audio content such as a points cloud a 3D mesh are not required to be marked.

(62) Digital twins that provide a virtual replica of persons, physical objects or systems, integrating multiple data types, also fall within the scope of Article 50(2) AI Act, if they are AI generated or manipulated and take the form of audio, image, video or text, unless they fall within the exceptions and qualify as ‘industrial applications’ or ‘business-to- business applications’ under point 87 below or real-time ephemeral content under point 88 below.

(63) Article 50(2) AI Act may also apply to AI agents if the AI agent takes an action, the output of which is AI-generated or manipulated content perceptible by natural persons in the form of audio, image, video or text. Those outputs must therefore be marked and detectable as described below. In contrast, any AI agent´s action that does not include AI-generated or manipulated content (e.g., intermediate processing steps such as reasoning and chain of thought), or AI-generated action that is not intended to be directly perceived by natural persons (e.g. a web request or browser action), does not qualify as synthetic content within the scope of Article 50(2) AI Act and does not need to comply with the marking and detection obligations.