AI disclosure rules change across the EU from 2 August

The roles

The provider and the user have different jobs

The Act separates providers from deployers.

A provider develops an AI system, or has one developed, then places it on the EU market or puts it into service under its own name. Providers of systems that interact directly with people must make the AI interaction clear unless it is already obvious. Providers of generative systems must also add machine readable marks that allow synthetic audio, images, video and text to be detected.

A deployer uses an AI system professionally under its authority. This can include a business, an agency, a studio or a freelancer earning regular income from the work. The Commission’s guidance says employees and contractors working under an organisation’s responsibility are not separate deployers. Responsibility stays with the organisation using the system.

Deployers have their own duties. They must disclose deepfakes, inform people when emotion recognition or biometric categorisation is being used, and label certain AI generated or manipulated text published to inform the public on matters of public interest.

Article 50 treats each use of AI according to the system, the output, its purpose and the audience that sees it. The presence of AI alone does not determine the disclosure.

The disclosure

A visible label and a hidden mark do different work

Machine readable marking is mainly a provider responsibility. It gives platforms and detection tools a way to identify the origin of synthetic content.

A deployer cannot rely on that hidden mark when a deepfake needs to be disclosed. The Commission says the disclosure must be clear and distinguishable when a person first encounters the content. It should be visible or audible without requiring a special tool.

The creative exception is more practical than some early summaries suggested. Where a deepfake forms part of an evidently artistic, creative, satirical or fictional work, disclosure is still required, but it can be made in a way that does not hamper the display or enjoyment of the work. The guidance also says that standard film effects or background manipulation are unlikely to count as deepfakes when an audience would not expect them to be authentic.

That leaves room for the work to remain the work. It also means the disclosure needs to be planned alongside the format, rather than added as an afterthought once the piece is finished.

The review

Human review has to reach the substance

The rules for text are narrower than a universal label on AI assisted writing.

They apply when AI generated or manipulated text is published to inform the public on a matter of public interest. The Commission’s examples include politics, public services, health, safety, the environment, and economic, financial, scientific or cultural developments that may be part of public debate.

There is an exemption where the text has undergone human review or editorial control and a person or organisation holds editorial responsibility for the publication.

The standard for review is meaningful. The Commission describes a deliberate examination of the substance by someone with relevant knowledge and professional expertise. An editor must be able to approve, change or reject the content, check facts and assess the sources. A spelling or grammar pass does not meet that standard.

That standard has value beyond compliance. If nobody can name the person who checked the claims, challenged the framing and accepted responsibility for publication, the workflow is unfinished.

The practical record

Build the record before you need the label

A small team can begin with a reliable record of how it uses AI.

Start with five questions:

• Where does a customer, employee or audience member interact directly with AI?
• Which published assets contain synthetic or materially manipulated audio, images or video?
• Could any of that work resemble a real person, place, object or event closely enough to appear authentic?
• Which public interest text uses AI, and who reviews its substance before publication?
• Where will the disclosure appear so that people encounter it clearly and accessibly?

Keep the original sources, the AI tool used, the changes made by a person, the name of the reviewer and the final approval. Preserve any provenance metadata supplied by the tool or platform instead of stripping it during export.

This record makes the questions visible early enough to act on them. Some cases will still need legal judgement.

For work that may fall within Article 50, check the current Commission guidance and take qualified legal advice on your specific circumstances.

A production record should show what AI did, what a person checked and who approved the work.