The Act does not require every piece of AI-assisted content to carry a label. Its transparency rules focus on specific uses, including deepfakes, public-interest text published without human review or editorial control, and direct interactions between people and AI systems.
The European Union’s Artificial Intelligence Act entered into force in 2024, but many of its transparency requirements began applying on 2 August 2026. For organisations that publish articles, campaign material, videos or social media content, the rules raise a practical question: when must the public be told that AI was used?
The answer depends on the type of content, the organisation’s role and the checks carried out before publication. Article 50 distinguishes between the responsibilities of AI system providers and those of organisations that use AI systems, known in the Act as “deployers”. The European Commission’s guidance on Article 50 sets out how the distinction applies.
A provider develops an AI system, or has one developed, and places it on the market or puts it into service. A deployer uses an AI system under its authority. An organisation using a generative AI tool to produce a public-facing image or article will generally be acting as a deployer; employees and contractors using it on the organisation’s instructions are not treated as separate deployers.
When content needs a label
The rules cover two main types of public-facing content. The first is a “deepfake”: an AI-generated or manipulated image, audio recording or video that resembles a real person, object, place, organisation or event and could falsely appear authentic. Organisations that publish such material must disclose that it has been artificially generated or manipulated. The disclosure must be clear and visible or audible to the audience; a technical marker hidden in the file is not enough.
The second is AI-generated or manipulated text published to inform the public about a matter of public interest. The Commission’s examples include political issues, public services, justice, fundamental rights, public health, environmental protection, consumer safety and developments likely to be relevant to public debate.
This does not mean every use of AI in drafting requires a label. Text that has undergone meaningful human review or editorial control is exempt from the labelling requirement. The reviewer must examine the substance of the material, while editorial control means a responsible editor or organisation has the authority to approve, change or reject it on substantive grounds, including checking facts and sources. A spellcheck or a quick grammar edit alone does not qualify.
That distinction is particularly relevant to organisations publishing policy explainers, campaign updates, research summaries and news material. If AI produces or materially changes the text, the organisation should be able to show what human review took place and who held responsibility for publication. If there was no substantive review or editorial control, the text must be clearly labelled when it falls within the public-interest category.
Chatbots and AI-generated media
The Act also addresses direct interactions between people and AI. Providers of systems such as chatbots, AI agents and avatars must ensure that users are told they are interacting with AI, unless this is obvious. The notice should appear at the start of the interaction. An organisation commissioning or deploying such a tool should check how that notice is presented to users.
Separately, providers of generative AI systems must mark their outputs in a machine-readable format so that AI-generated or manipulated content can be detected. That technical obligation rests with providers. It does not replace the deployer’s duty to give people a clear disclosure when publishing a deepfake.
What communications teams should do
Organisations can start by listing where they use AI: website copy, press materials, social posts, campaign graphics, video, audio and customer-facing chat tools. For each use, they should identify whether the material is generated or materially manipulated, whether it concerns a matter of public interest, and what review is required before publication.
A workable editorial process should record who checked factual claims, assessed sources and approved the final material. Where a label is required, it should be placed where audiences will see or hear it, such as alongside an image or at the start of a video. The Commission’s questions and answers on Article 50 explain that disclosures for deepfakes must be understandable without special tools or extra effort.
The rules do not require organisations to abandon AI tools. They require clearer responsibility for how certain AI-generated material reaches the public. For communications teams, the practical test is whether the content could be mistaken for authentic material, whether it informs the public on an issue of public interest, and whether a person with relevant knowledge has checked and taken responsibility for it.
