EU AI Act Introduces New Labelling Rules for Deepfakes and AI-Generated Content

Summary

• Article 50 transparency obligations apply from 2 August 2026
• Deepfakes must carry a clear disclosure when used professionally
• Generative AI providers must incorporate machine-readable markings
• Chatbots must inform users that they are interacting with artificial intelligence
• Special provisions apply to artistic, satirical and fictional works
• Content created before the rules take effect does not require retroactive labelling

From 2 August 2026, new transparency obligations will apply across the European Union to deepfakes, conversational systems and specific categories of content generated or manipulated with artificial intelligence.

The rules form part of Article 50 of the EU AI Act and require providers and professional deployers of artificial intelligence systems to inform the public when people are interacting with AI or viewing certain forms of synthetic content.

The development is particularly significant for photographers, visual creators, news organisations and online platforms, as it seeks to make the distinction between genuine captures, edited material and images produced entirely by generative artificial intelligence clearer.

What applies from 2 August 2026

Article 50 of the AI Act applies from 2 August 2026 and introduces different obligations depending on whether an entity provides an AI system or uses it in a professional context.

Providers of systems that generate or manipulate images, audio, video and text must design their services so that their output is marked in a machine-readable format. The aim is to make it technically possible to detect that content has been generated or altered by artificial intelligence.

At the same time, anyone professionally deploying an AI system to create a deepfake must disclose that the visual or audio material has been artificially generated or manipulated.

The new requirements do not mean that every file involving any form of AI tool must automatically display the same visible label. The framework distinguishes between technical marking obligations for providers and public disclosure requirements applying to specific uses, including deepfakes and certain public-interest publications.

What is considered a deepfake

Under the AI Act, a deepfake is image, audio or video content generated or manipulated by AI that resembles existing people, objects, places, entities or events and could falsely appear authentic.

The disclosure requirement is intended to inform viewers in good time that what they are seeing or hearing is not an unaltered record of reality.

For a photographer or video creator, this may include a synthetic depiction of a real person, a scene presented as an actual event that never occurred or a video in which a person’s appearance, voice or actions have been digitally altered.

Obligations for chatbots and interactive systems

The rules are not limited to images. Systems that interact directly with people, including chatbots and digital assistants, must clearly inform users that they are communicating with artificial intelligence.

An exception may apply when it is already obvious from the circumstances and the way the service operates that the interaction involves an AI system.

This requirement seeks to prevent situations in which users believe they are communicating with a real employee, company representative or other natural person.

What applies to public-interest text

A specific provision covers text generated or manipulated with AI and published for the purpose of informing the public about matters of public interest.

In such cases, the use of artificial intelligence must be disclosed unless the content has undergone human review or editorial control and a natural or legal person assumes editorial responsibility for its publication.

This distinction is important for publishers and news organisations. Using an AI tool as a supporting resource is not automatically treated in the same way as publishing autonomous, unchecked content, provided meaningful human oversight and clear editorial responsibility are present.

Exceptions for artistic and satirical works

The European framework provides special treatment for works forming part of evidently artistic, creative, satirical, fictional or comparable content.

The disclosure obligation still exists, but it must be applied in a way that does not hamper the display, enjoyment or coherence of the work. This allows the information to be adapted to the nature and format of the creation without undermining the artistic experience.

The exception is not a blanket exemption for every image described as art. The objective remains to inform audiences when synthetic material could be mistaken for an authentic record.

No retroactive labelling requirement

Content generated before 2 August 2026 does not have to be labelled retroactively. The European Commission nevertheless encourages voluntary disclosure where it is feasible and useful.

A limited transition period applies to certain generative AI systems that were already placed on the market before the date of application, specifically in relation to technical marking and detection requirements.

This does not remove the remaining transparency obligations taking effect on 2 August, but recognises that integrating machine-readable markings into existing systems may require technical modifications.

How labels may be displayed

Information provided to the public must be clear and distinguishable. Depending on the type of content, it may appear as a visual notice, caption, label, announcement before playback or another prominent form of disclosure.

At file level, providers may use technical solutions such as metadata, digital provenance credentials, watermarks or other machine-readable signals. To the extent technically feasible, these markings should be effective, interoperable, robust and reliable.

The European Union has also introduced common icons that may be used to disclose deepfakes and AI-generated text, helping audiences become familiar with a more consistent visual language.

Concerns about implementation

Representatives of the technology industry have raised concerns that the use of multiple different labels could create confusion rather than provide clearer information.

Google’s Karen Massin argued that an environment filled with overlapping labels and legal disclosures could make it more difficult for users to understand the actual context of each publication.

Ashley Casovan of the International Association of Privacy Professionals responded that implementation difficulties are a common reaction to new compliance requirements and that organisations eventually develop the processes needed to meet them.

What this means for photographers and creators

For the photography industry, the rules increase the importance of documenting the origin and editing history of an image.

Routine adjustments to exposure, colour or noise do not necessarily amount to the creation of deceptive synthetic content. However, the more substantially an edit alters people, events or locations to create a convincing but nonexistent reality, the more likely it becomes that disclosure requirements may apply.

Professionals will need to consider not only which tool they used but also the final result, the way it is presented, its intended audience and whether there is a risk of deception.

The development may also strengthen provenance technologies that allow viewers to check when and with which device a file was created and which changes were made later.

Potential penalties

Failure to comply with transparency obligations may lead to administrative fines. For certain violations, the AI Act sets a maximum of up to €15 million or, for companies, up to 3% of their total worldwide annual turnover for the preceding financial year, whichever is higher.

Enforcement and the precise amount of a penalty depend on the nature, seriousness and duration of the infringement, as well as factors including the size of the organisation and its level of cooperation with the competent authorities.

What we think

Mandatory transparency is a necessary step at a time when synthetic images are becoming increasingly difficult to distinguish from photographs. The effectiveness of the rules, however, will depend on labels being simple and consistent, remaining attached during redistribution and being meaningfully enforced by platforms and professionals. For photographers, reliable documentation of an image’s origin could become an increasingly important element of professional credibility.

Frequently asked questions

Does every image edited with AI need to be labelled?

Not necessarily. The obligations depend on the type of editing, the role of the provider or professional deployer and whether the result constitutes a deepfake or another category of content covered by Article 50.

Do the rules apply to older AI images?

Mandatory retroactive labelling does not apply to content created before 2 August 2026, although voluntary disclosure is encouraged.

Are artistic works covered?

Special provisions apply to artistic, creative, satirical and fictional content. Disclosure should be made in a way that does not hamper the display or enjoyment of the work.

What must chatbots disclose?

Users must be informed that they are interacting with an artificial intelligence system unless this is already obvious from the circumstances.

Are image metadata sufficient?

Machine-readable marking primarily concerns provider obligations. When public disclosure is required for a deepfake or other covered content, the information must be clear and accessible to the audience rather than hidden exclusively in metadata.