Deepfake label: how to disclose a deepfake under the AI Act

The AI Act (Regulation (EU) 2024/1689) writes it in two words, deep fake. Article 3(60) defines it as “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”. In plain words: a picture made or changed with AI that looks real, and that someone could take for real.

The Commission's Guidelines on Article 50 (paragraph 113) split this into four criteria. All four must be met:

1. Resemblance: a high level of similarity to the subject, judged on its distinctive features. It doesn't have to be identical.
2. Existing: the subject exists, or could plausibly exist or have existed. A photorealistic person who never lived can qualify; a dragon or an elephant driving a car can't.
3. Persons, objects, places, entities or events: including realistic AI avatars, buildings, artworks, products, animals and realistic scenes.
4. False appearance: the content would falsely appear authentic or truthful to someone.

The Guidelines give examples on both sides of the line. Some are videos, but the reasoning is the same for a picture.

• No intent needed: the test is objective (paragraph 114). An image can be a deep fake even if you never meant to deceive anyone.
• Context and audience count: where the image is shown, and what the people likely to see it expect, including children or people less familiar with AI (paragraph 115). If part of that audience could be misled, that can be enough.
• Photorealism makes a deep fake more likely, but isn't decisive on its own.
• Minor edits usually don't count (paragraph 116): removing passers-by, light and colour correction, noise reduction, cosmetic adjustments, or an aesthetic background change in a product ad. Substantial AI editing of a news photo's background may.

Article 50(4) puts the duty on deployers: people and organisations using AI in a professional capacity. Business, trade and freelance work count, and so does gaining an economic benefit on a regular basis (Guidelines, paragraph 19). Staff act for their employer: the company is the deployer (paragraph 14).

Purely personal use is excluded (Article 2(10)), such as deep fakes of your household or ones you post on social media in a personal capacity. Platforms may still ask you to label: see AI labels on social media.

• Since 2 August 2026, with no grace period for deployers. The Digital Omnibus delayed only the providers' machine-readable marking, to 2 December 2026, for generative AI systems placed on the market before 2 August 2026.
• No retroactive duty for deep fakes made before 2 August 2026, though labelling them is encouraged (paragraph 154).

Fines and enforcement are in our guide to the EU AI labelling law.

Article 50(4) asks deployers to disclose that the content has been artificially generated or manipulated. Article 50(5) adds “in a clear and distinguishable manner at the latest at the time of the first interaction or exposure”, and in an accessible way. For an image, that means:

• Clear and distinguishable: noticeable and easy to tell apart from its surroundings, not buried in a menu or the terms of use (Guidelines, paragraph 142).
• At first exposure: for every person who sees it, for example while scrolling a feed (paragraph 143).
• Visible to people: the AI provider's hidden watermark or metadata doesn't meet the deployer's duty (paragraph 117). Our guide to Detect AI-generated images explains what those marks do.
• Embedded in the image, visible against any background, where no other overlay gets in the way, “e.g., in the top right corner” (Code of Practice, Measure 1.2).
• Accessible: high contrast, and alt text for screen readers.

The Code accepts the EU icon or an equivalent label. The icon is optional, and using it doesn't by itself establish compliance.

Pick the European Commission's icon that says what happened to the picture:

For a face swap in a real photo, use AI MODIFIED: a face in an authentic photo swapped for a politician's face is the Commission's own example. The Code also encourages saying what changed, such as “a face has been altered”. Colours, sizes and downloads are in our guide to the EU AI icons.

For a deep fake in an evidently artistic, creative, satirical, fictional or analogous work, Article 50(4) still requires a disclosure, but one that doesn't hamper the display or enjoyment of the work. The Guidelines read “evidently” strictly (paragraph 122): the work's nature must be obvious from its style, its context or what the audience expects. Ambiguous content gets the standard label, and so does purely informative or commercial content, such as news. Where informative and creative sides mix, the informative side prevails.

The Code (Commitment 3) suggests disclosing in the notes, the description or the credits, or next to the image frame on a website, as long as people notice it without clicking. On the picture itself, a smaller label or a quieter corner may suit the work, if it stays easy to see.

A label meets one AI Act duty and nothing more: complying with transparency rules doesn't make an AI use or its output lawful (recital 137). You still need the consent of the people shown, or of the rightsholders, where the law requires it (Code, Commitment 1). Data protection, copyright, trade marks, personality rights over a person's image and voice, misleading-advertising rules and criminal law all still apply (Guidelines, paragraphs 127 to 129).

The tool adds the icon; whether and how to label stays your decision and your responsibility.