What is a deep fake under the EU AI Act?

The legal definition

“‘deep fake’ means 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;”

Article 3(60), Regulation (EU) 2024/1689

Recital 134 adds that the resemblance has to be “appreciable”. The Commission's guidelines on Article 50, C(2026) 5054, examine the definition through four criteria that must all be fulfilled (para. 113). AI also has to be involved; a retouch done without AI is outside the definition.

The four criteria

1. Resemblance

The content must show a high level of similarity to its subject, including any recognisable elements, though it need not be identical. Whether the resemblance is appreciable is for the deployer to assess case by case, through an objective comparison that looks at how far characteristic or distinctive features are reproduced (para. 113 i).

2. Existing or plausibly existing

It is enough for the person, object or scene to exist, to be able to plausibly exist or to have plausibly existed (para. 113 ii), because according to the guidelines such content can be a major source of deception and disinformation. What defies the laws of nature or physics, or shows lifeforms not accepted in biology, and has no potential to mislead, is out of scope, such as humans flying without mechanical aids, dragons and elephants driving cars.

3. Persons, objects, places, entities or events

The guidelines describe the five categories as follows (para. 113 iii):

• Persons: realistic human beings, including digital replicas of real people, AI-generated avatars or personas, and personal characteristics such as image, voice and behaviour.
• Objects: realistic, inanimate material items such as buildings, artworks, machinery and consumer goods.
• Places: realistic locations.
• Entities: realistic, non-human but animate beings such as animals and other lifeforms.
• Events: realistic scenes or situations involving any of the above, such as historical events or the depiction of services.

A photorealistic AI image showing a product, a person, an animal or a room therefore nearly always meets the first three criteria. In an online store, the decision usually turns on the fourth.

4. False appearance of authenticity or truthfulness

The guidelines separate the two ideas (para. 113 iv). Authenticity concerns source and creation, meaning whether real people or animals were involved, whether an object looks and is used like that and whether an event unfolded that way. Truthfulness concerns whether what is shown is factually correct. This criterion is to be understood independently of the notion of deception in the Unfair Commercial Practices Directive (guidelines, footnote 32). Arguing that an informed consumer would spot the AI is therefore beside the point.

An objective test

Whether content falsely appears authentic is assessed as a whole, looking at the level of resemblance, the message the content may carry, the intended and foreseeable context, the environment in which it is shown, and who the audience is and what it expects (para. 114). The assessment is objective; no intention to deceive or mislead is required.

A high degree of photorealism makes a deep fake more likely, but it is not decisive on its own (para. 114). If the audience does not expect real footage, context can rule out a false appearance; the guidelines point to AI backgrounds and special effects in standard film production. Fully AI-generated actors, digital replicas and de-aging, by contrast, are likely to mislead.

The broadest foreseeable audience

Unlike Article 50(1), the test is not based on a hypothetical average person but on how varied the foreseeable audience may be (para. 115), especially where children, older people or people with low digital and AI literacy are likely to see the content. If it would falsely appear authentic to that part of the audience, that can be enough.

The limit is the reasonably foreseeable audience. Deployers do not have to anticipate onward sharing by third parties beyond their intended channels, for example for content shown only to subscribers or in a corporate newsletter (para. 115). A publicly accessible online store is open to everyone.

Minor changes

AI changes to insignificant substantive or technical aspects of existing content generally do not turn it into a deep fake (para. 116). The guidelines list editing background details such as removing passers-by, lighting adjustments, colour correction, noise reduction, accessibility improvements, compression and cosmetic enhancements. In advertising and on packaging, colour correction, extending existing content, replacing backgrounds for clearly aesthetic purposes, arranging existing products and rescaling are likely to have only a minor effect.

According to the guidelines, context and effect in the specific case decide. In journalistic images, substantial AI editing of background details can already affect how authentic the picture seems.

Examples from the guidelines

These examples from the guidelines are the most relevant for retailers:

The car example marks the boundary. An AI background does not make the image a deep fake as long as it is only scenery. Once the scene says something untrue about the product, for instance about its size, where it can be used or what it does, the answer changes. More store situations are collected under Cases.

How this differs from Article 50(2)

Article 50(2) and 50(4) are separate obligations (para. 111). Paragraph 2 requires providers of generative AI systems to mark their synthetic output in a machine-readable format whether or not it is a deep fake, so text and fantasy images too. Systems that only assist with standard editing or do not substantially alter the input are exempt (paras. 90 to 92). Paragraph 4 applies to deployers and only to deep fakes, and it requires a visible disclosure. The standard editing list in para. 92 formally belongs to paragraph 2; for paragraph 4, the deep fake definition and para. 116 are what count.

Deployers cannot rely on the provider's marking, because people cannot perceive it (para. 117). Why it is often missing is covered in AI metadata; what a deep fake means for labelling is explained in the labelling obligation.