The transparency obligations under Article 50 of the European AI Act have applied since 2 August 2026. Their purpose is to help people recognise when they are interacting with an AI system or encountering certain artificially generated or manipulated content. One distinction is particularly important for companies, agencies, editorial teams and website operators: the AI Act does not impose a blanket requirement for every AI-generated image on every website to carry a visible label.
The Regulation assigns different responsibilities to providers and deployers of AI systems. A company that develops a generative AI system or places it on the European market under its own name may be a provider. A company that uses such a system under its own authority in a professional or commercial context will generally be a deployer. Employees, designers or commissioned freelancers do not automatically become separate deployers when they act under a company’s authority.
When did Article 50 of the EU AI Act start to apply?
Article 50 has applied since 2 August 2026. This is confirmed by both the official EU AI Act Service Desk and the European Commission’s guidelines on the transparency obligations. From that date, providers and deployers within the scope of the rules must comply with the requirements that apply to them.
According to the Commission’s current information, a limited transitional arrangement applies only to certain generative AI systems that were placed on the market or put into service before 2 August 2026. Their technical marking and detection obligation under Article 50(2) is subject to a transitional period until 2 December 2026. This specific arrangement is not a general postponement of visible deepfake disclosures or other obligations for deployers.
Technical marking: what providers of generative AI must do
Providers of AI systems that generate synthetic audio, image, video or text content must design their systems so that the outputs are marked in a machine-readable format and can be detected as artificially generated or manipulated. The technical solution must be effective, interoperable, robust and reliable as far as technically feasible. The characteristics of the medium, the state of the art and implementation costs may be taken into account.
This obligation is primarily directed at providers of generative systems. A company that creates an image using a third-party image generator and then publishes it on its website does not automatically have to develop the technical provenance marking itself. However, it should avoid unnecessarily removing existing metadata, provenance information or other machine-readable features and should review whether its export and optimisation processes preserve them.
An exception applies where the AI system merely performs an assistive function for standard editing or does not substantially alter the input data or its meaning. Not every automatic correction, noise reduction process or standard retouching operation therefore turns content into synthetic material that must be labelled.
Visible disclosure: the definition of a deepfake is decisive
For professional deployers, Article 50(4) is particularly relevant to images, audio and video. Content classified as a deepfake must be visibly disclosed. Under the AI Act, a deepfake is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.
According to the European Commission’s current questions and answers on Article 50, three characteristics must be present:
- Resemblance: The content closely resembles the person, object, place, entity or event depicted.
- Existence or plausible existence: The simulated person, place, object or event exists, could plausibly exist or could plausibly have existed.
- False impression of authenticity: The content could deceive or mislead people about whether it is authentic or truthful.
This means that not every abstract AI motif, obviously fantastical illustration or stylised visual world is automatically a deepfake that requires disclosure. Conversely, a photorealistic image may be problematic if it appears to be an authentic photograph of a real person, location, product or event. Relevant factors include the subject, degree of resemblance, message, context of use and audience expectations.
How must a deepfake disclosure be presented?
The disclosure must be clear and distinguishable no later than the point of first exposure. It must be understandable and perceivable without special tools, for example as a visible label on an image or video or as an audible notice in audio content. An embedded machine-readable marker on its own is not sufficient for the deployer’s disclosure obligation.
The label must also comply with applicable accessibility requirements. On a website, a notice should therefore not rely on colour alone. A practical solution combines legible text, sufficient contrast, reliable positioning within the medium and an additional machine-readable description for assistive technologies.
The EU provides optional icons for labelling AI-generated content. According to the Commission, the icons performed better in user testing when accompanied by a clear text description. The icons are a supporting tool; the decisive point is that the specific disclosure is clear, distinguishable and appropriate for the medium.
Creative, artistic and fictional content
The AI Act provides an adapted rule for content that is evidently part of an artistic, creative, satirical, fictional or comparable work or programme. The disclosure may be presented in a way that does not hamper the display or enjoyment of the work. This is not a complete exemption, but it permits an appropriate and less intrusive form of labelling.
On a designed website, this could mean a discreet label within the image, a notice in the caption or clearly associated information placed immediately alongside it. Whether a particular solution is sufficient depends on the individual case and on whether the image could be understood as a potentially authentic depiction.
What applies to AI-generated text?
Article 50(4) sets a narrower threshold for text. Disclosure is required where an AI system generates or manipulates text that is published for the purpose of informing the public on matters of public interest. The Commission identifies areas including politics, public administration, the administration of justice, fundamental rights, public safety and health, environmental and consumer protection, as well as relevant economic, scientific and cultural developments.
The disclosure obligation does not apply where the text has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility for its publication. According to the guidelines, a simple spelling check, formal approval or superficial review is not sufficient. The content must be assessed by a competent reviewer who can amend or reject statements and verify the sources.
A practical process for websites and content teams
- Clarify roles: Document which AI systems are used and whether the company acts as a provider, a deployer or both.
- Inventory content: Record AI-generated and substantially AI-manipulated images, videos, audio and text in a media inventory.
- Assess deepfake risk: For realistic content, document whether it depicts real or plausibly real people, places, objects or events and could appear authentic.
- Preserve technical markings: Check whether image optimisation, the CMS, social media exports or the CDN remove existing provenance data or machine-readable markings.
- Assign visible disclosures: Place labels directly on the relevant medium, keep them within the visible area when responsive cropping is used and provide an accessible description.
- Document text review: For publications on matters of public interest, record who performed the substantive review, checked the sources and assumed editorial responsibility.
- Explain exceptions: Where content is treated as standard editing, an unmistakably creative depiction or otherwise outside the disclosure requirement, document the decision briefly and clearly.
- Prioritise existing content: According to the Commission, deepfakes generated and published before 2 August 2026 do not have to be labelled retrospectively, although voluntary disclosure is encouraged. Newly published or newly generated content should be included in the process from the outset.
Voluntary EU Code of Practice and potential penalties
The EU Code of Practice on the transparency of AI-generated content is voluntary. The Commission and the European Artificial Intelligence Board have assessed it as an appropriate instrument for demonstrating compliance with Article 50(2), (4) and (5). Organisations that do not sign the Code remain bound by the legal obligations and may need to demonstrate by other means that their measures are appropriate.
According to the European Commission’s current quick facts on transparency rules for AI systems, infringements of other obligations under the AI Act may be subject to fines of up to EUR 15 million or, for undertakings, up to three per cent of worldwide annual turnover. The statutory proportionality requirements apply when penalties are determined, including specific consideration for small and medium-sized enterprises.
