Deepfake & AI Content Labelling Under the EU AI Act (Article 50(4))

Deepfake & AI content labelling under Article 50(4) of the EU AI Act is one of the most broadly applicable obligations from 2 August 2026. Anyone who uses AI to create realistic synthetic media for an EU audience faces a direct duty: label it. Article 50(4) of the EU AI Act is the deepfake and AI-content labelling obligation, and it lands on a far wider group than “deepfake” suggests — marketing teams, publishers, agencies, public authorities and educators all fall inside it the moment they publish AI-generated images, audio, video or certain text.

This guide explains what counts as a deepfake under the law, the two things Article 50(4) actually requires, the artistic and editorial exemptions and their hard limits, how to label correctly, and what to do before the deadline.

Deepfake AI content labelling: what Article 50(4) requires

Article 50(4) of Regulation (EU) 2024/1689 places a labelling duty on deployers — the organisations that actually use an AI system to create and publish content. It covers two distinct categories:

• Deepfakes. Deployers of an AI system that generates or manipulates image, audio or video content constituting a deepfake must disclose that the content has been artificially generated or manipulated.
• AI-generated public-interest text. Deployers who publish AI-generated or AI-manipulated text to inform the public on matters of public interest must disclose its artificial origin.

The purpose is to counter the manipulation risk that realistic synthetic media carries. The duty sits on the deployer because the deployer controls the act of publication — the moment the content reaches the public.

What counts as a “deepfake”

The definition is in Article 3(60): 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. The draft guidelines sharpen this into a practical test. Content is a deepfake where it:

• resembles real persons, objects, places, institutions or events — including realistic-looking but fictional subjects; and
• could be mistakenly perceived by a person as genuine or truthful.

The key boundary: clearly fantastical or physically impossible content falls outside the definition. An AI image of a dragon, or of a human flying unaided, is not a deepfake — no reasonable person would take it as real. A photorealistic image of a person who does not exist, but looks entirely real, is covered.

One point that surprises people: the labelling duty does not depend on any intention to deceive. If the content is realistic and could be mistaken for genuine, it must be labelled — even where it was made in good faith and no deception was intended.

The exemptions — and their hard limits

Article 50(4) contains real carve-outs, but each is narrower than it first appears.

Artistic, creative, satirical or fictional works — attenuated, not exempt

Where a deepfake forms part of an evidently artistic, creative, satirical, fictional or analogous work or programme, the obligation is reduced, not removed. Disclosure is limited to revealing the existence of the generated content in a manner that does not hamper the display or enjoyment of the work — for example, in the credits or an unobtrusive note rather than a label stamped across every frame.

The hard limit: this attenuation does not rescue realistic depictions of real people in misleading ways. A clearly fictional AI film is fine; an AI-generated video of a real politician saying things they never said is not, regardless of how it is framed. Third-party rights — intellectual property, personality and privacy rights — must still be respected.

Editorial responsibility — for public-interest text only

The public-interest-text duty does not apply where the content has undergone genuine human review or editorial control and a natural or legal person holds editorial responsibility for the publication. But the Code of Practice narrows this sharply: it requires a documented editorial workflow with identified responsible persons, not a mere assertion that “a human looked at it.” A cursory skim does not qualify.

Law enforcement

The labelling duty does not apply where the use is authorised by law to detect, prevent, investigate or prosecute criminal offences.

Check Article 5 first — a sequencing trap

Before reaching the Article 50 transparency question, check whether the use is outright prohibited under Article 5. Some AI practices are banned entirely — if a system falls under an Article 5 prohibition, the transparency analysis is moot because the system cannot lawfully operate at all. Transparency is the rule for permitted AI; it never legitimises a prohibited use.

How to label correctly

The regulation does not mandate one fixed format, but the guidelines and Code point to clear, perceivable methods, and to what fails. The proposed approaches include:

• Persistent visual labels on images and video — visible markers that travel with the content.
• Opening disclaimers at the start of video content.
• Audible warnings for audio content.
• Clear, perceivable disclosure for public-interest text, at the point the reader encounters it.

What does not satisfy the duty: a machine-readable watermark on its own (it is not perceivable by an ordinary viewer), a disclosure buried in terms and conditions, or a vague note that “AI tools were used.” The disclosure has to be perceivable in the content itself. A standardised EU label is being developed to support consistent practice.

Note the interaction with Article 50(2): even where a deployer’s labelling is attenuated or exempt, the provider of the generation system may still owe the separate machine-readable marking duty under 50(2). The two operate together — perceivable labelling for humans, machine-readable marking for detection tools.

The Code of Practice — voluntary, but the benchmark

The Code of Practice on Transparency of AI-Generated Content supports the marking and labelling workflows for Article 50. It is voluntary, and not legally binding — but it functions as the recognised compliance benchmark. Organisations that adhere to it provide strong evidence of compliance; those that ignore it carry the full risk and must justify their alternative measures. Notably, the Code reframes labelling as a built-in product feature: it anticipates interface-level options that embed perceptible deepfake labels at the point of generation, free detection tools with confidence scores, and an audit-ready compliance record that may be shared with market surveillance authorities.

Penalties

Breach of the Article 50 transparency obligations sits in the AI Act’s middle penalty tier: up to €15 million or 3% of total worldwide annual turnover, whichever is higher. For an organisation that publishes synthetic media at scale, this is not a theoretical risk.

A labelling compliance checklist

• Map externally published AI content — website, social media, ads, reports, video, audio. You cannot label what you have not located.
• Classify each item. Deepfake (image/audio/video resembling real or realistic subjects)? Public-interest text? Or outside scope?
• Check Article 5 first for any prohibited use, then apply Article 50(4).
• Plan disclosure from the outset for synthetic media — persistent labels, opening disclaimers, audible warnings.
• Test your exemptions honestly. “Evidently artistic” must be evident; “editorial responsibility” needs a documented workflow with named owners.
• Coordinate with providers on the 50(2) machine-readable marking so human-facing and machine-readable disclosure align.
• Keep an audit-ready record of what you labelled, how, and why — the evidence a market surveillance authority may request.

Frequently asked questions

Is AI-generated art exempt from labelling?

Not fully. Evidently artistic, creative, satirical or fictional works get an attenuated duty — disclosure in a way that does not spoil the work (e.g. credits) — not a full exemption. And it never covers realistic depictions of real people in misleading ways.

Do I have to label an obviously fake image, like a dragon?

No. Clearly fantastical or physically impossible content falls outside the deepfake definition. The duty targets content that could be mistaken for real.

What if I didn’t intend to deceive anyone?

Intent is irrelevant. If the content is realistic and could be perceived as genuine, the labelling duty applies regardless of good faith.

Does labelling AI-written articles count here?

Only where the text is published to inform the public on matters of public interest — and even then, a documented editorial-responsibility workflow can remove the duty. Ordinary marketing copy is generally outside 50(4), though provider-side 50(2) marking may still apply.

Is a watermark enough?

A machine-readable watermark alone does not satisfy the deployer’s labelling duty, because it is not perceivable by an ordinary viewer. You need a human-perceivable label as well.

When does this apply?

2 August 2026 under the law in force as of June 2026. The Digital Omnibus reform leaves Article 50 on this schedule; confirm the current position before relying on the date.

The bottom line

Article 50(4) turns “we made this with AI” from a nice-to-have disclosure into a legal duty for realistic synthetic media aimed at the public. If you publish AI-generated images, audio, video, or public-interest text to an EU audience, plan the label from the moment of creation — not as an afterthought. The exemptions are real but narrow, the penalties are serious, and the organisations that build labelling into their content pipeline now will not be scrambling on 2 August.

Related guides: Article 50 transparency overview · chatbot disclosure · compliance timeline · Commission AI Act resources.