The European Commission's guidelines on Article 50 of the EU AI Act (C(2026) 5054 final, content approved on 20 July 2026) explain when an AI image is a deep fake, who has to label it and how. They are not binding: only the Court of Justice of the EU can interpret the Regulation authoritatively (para. 5).
Authorities still follow them. For stores, the key passages are the deployer role (paras. 12 to 19), the deep fake definition (paras. 111 to 129), form and timing (paras. 142, 143), evidence (para. 148) and the transition rules (paras. 153, 154).
Key facts
- Reference: C(2026) 5054 final. The Commission approved the content on 20 July 2026. According to its cover document, formal adoption follows once all language versions are available.
- The guidelines are not binding. Only the Court of Justice of the EU can give an authoritative interpretation (para. 5).
- A retailer is a deployer when it decides to use AI, including when freelancers or agencies work on its instructions (paras. 12, 14).
- Consumer goods can be the subject of a deep fake. It is enough that something could plausibly exist (para. 113).
- A notice only in manuals, menus or terms of use is not enough (para. 142). The AI provider's machine-readable marking does not count as a label (para. 117).
- Deep fakes generated before 2 August 2026 need no retroactive label, although the guidelines encourage one (para. 154).
What the guidelines are, and what they are not
Article 50 of the EU AI Act (Regulation (EU) 2024/1689) states the transparency duties in a few sentences. The European Commission's reading of them is set out in its guidelines on the transparency obligations of providers and deployers, reference C(2026) 5054 final. The Commission approved the content on 20 July 2026. Its cover document says the guidelines will be formally adopted once all language versions exist. We have not been able to confirm that this has happened. Paragraph numbers could shift at that point, so this page cites the version of 20 July 2026.
The guidelines are explicit about their own status: “These Guidelines are non-binding.” Only the Court of Justice of the European Union can interpret the Regulation with binding effect (para. 5). In practice they are still the main yardstick, because market surveillance authorities apply them. In Germany that is the Federal Network Agency (Bundesnetzagentur). Where the guidelines consider a label unnecessary, this site therefore says “not required under the guidelines” and cites the paragraph.
A separate document is the Code of Practice on Transparency of AI-Generated Content (final version of 10 June 2026). It is a voluntary code with concrete design rules, such as the “AI” mark. The guidelines interpret the law; the Code describes one way of meeting it. See Code of Practice.
Structure: the parts retailers should read
The guidelines run to about 50 pages and follow the structure of Article 50. The sections on chatbots (paragraph 1) and emotion recognition (paragraph 3) do not affect product images. Five parts matter for an online store:
- Introduction and status (including paras. 2, 5, 9): application from 2 August 2026, non-binding nature, protective aims such as preventing deception.
- Roles (including paras. 11 to 19): who is a provider, who is a deployer, and who merely passes content on.
- Machine-readable marking under Article 50(2) (including paras. 54 to 98): duties of AI providers and the list of standard editing (para. 92). Background knowledge for retailers.
- Deep fakes and text under Article 50(4) (paras. 111 to 138): the core for stores.
- Form, timing, enforcement, transition (paras. 142 to 154): what a label must look like, what you should be able to show without signing the Code, and what applies to older images.
The most practical parts are the example boxes after paras. 92, 116 and 124. They classify concrete cases as deep fake or not, and as a creative work or not.
The paragraphs that matter for images
The table sums up each paragraph in a sentence or two. Quotations are from the version of 20 July 2026.
| Para. | What it says | What it means for your store |
|---|---|---|
| 12 | The deployer is whoever decides to use an AI system and how. Technical control is not required. In longer chains, the deployer should take proportionate steps so that its label reaches the audience, for example through contracts with distribution partners. | If you decide to use an AI tool, you are the deployer. When you pass images to marketplaces or partners, make sure your label survives. |
| 14 | Employees are not separate deployers. A company that uses freelancers or contractors under its responsibility remains the deployer. A company that merely commissions an agency without deciding on its use of AI is not. | Agency images made on your instructions count as yours. Put in writing who decides about AI. |
| 16 | Actors who only disseminate third-party AI content are not deployers, but they are strongly encouraged to preserve existing markings and labels. | Manufacturer and supplier images: never strip existing labels or metadata. |
| 17 | Labelling content, or using labelled content, does not by itself make someone a deployer. Other actors are encouraged to inform their audience proportionately. | A voluntary label on a supplier image does not change your role. Where the AI origin is known, a label is recommended. |
| 19 | Any activity that brings a regular economic benefit is professional. Only purely personal use is excluded. | There is no private-use exception for images in a store. |
| 92 | Example list of standard editing (minor cropping, minor colour corrections, removing dust spots, removing backgrounds visible in the original file, edge completion) and of substantive changes (inserting or removing objects or persons, changing a face or body shape, extreme colour changes). | The list belongs to the provider exception in paragraph 2. When classifying your own images, read it together with para. 116. |
| 111 | Article 50(4) contains two separate duties (deep fakes and text). They apply in addition to the marking under Article 50(2). | A metadata entry written by the AI tool does not replace your visible label. |
| 113 | Four criteria: resemblance; existing or plausibly existing; persons, objects, places, entities or events; falsely appearing authentic or truthful. “Objects” expressly includes “consumer goods”. | Your product itself can be the subject of a deep fake. An invented but believable model counts as well. |
| 114 | The assessment is holistic and objective. No intention to deceive is required. Photorealism makes a deep fake more likely but is not decisive on its own. | “I did not mean to mislead anyone” is no reason to skip the label. |
| 115 | The yardstick is the foreseeable audience in all its diversity, including children, older people and people with little AI literacy. Deployers need not plan for further sharing by third parties. | A public store reaches everyone. “My customers can spot AI” does not hold. |
| 116 | Insignificant changes do not create a deep fake, for example colour correction, backgrounds changed for clearly aesthetic purposes, or arrangements of existing products. Deep fake: an AI product image that makes the product look different, more appealing or of higher quality than in reality. Not a deep fake: a real car against an AI background, as long as the ad does not mislead about the product. | The key paragraph for product images. A real, unchanged product on a decorative background: generally not required under the guidelines. A scene that says something about size, use or effect: check carefully. |
| 117 | Labelling must be perceivable without technical tools and without any action by the viewer. Deployers cannot rely on the provider's machine-readable marking. | The label must be visible on the image, with no click and no hover. |
| 119 | Evidently artistic, creative, satirical or fictional works benefit from a lighter form of disclosure (Article 50(4), third sentence). | Relevant for art prints you sell, not for product advertising. |
| 122 | The work categories are to be read narrowly. Content that is purely informative or commercial falls outside them. Where purposes are mixed, the informative one prevails. | An advertising image is generally not a work in the sense of the exception. |
| 123 | Deep fakes within works still have to be disclosed, only in an adapted form. | An art print still needs a notice. |
| 124 | Third-party rights such as data protection and intellectual property are unaffected. The example box treats teleshopping with simulated people and a synthetic influencer testing a product as not being works. | Changing real models with AI generally requires their consent. The label does not replace it. |
| 129 | The labelling duty does not permit deep fakes that are unlawful under other law, such as misleading advertising. | A label does not make an embellished product image lawful. |
| 142 | Clear and distinguishable means easy to identify as separate information and not easily overlooked. Not enough: only in manuals, menus or terms of use. | A notice in your terms, legal notice or an info page is not enough on its own. |
| 143 | First exposure applies to every person and every output, meaning any moment at which someone is reasonably likely to see the content. | Labels belong on product cards, search results, banners and blog posts, not just on the product page. |
| 148 | Those who do not sign the Code should show by other adequate means how they meet their duties, for example through a gap analysis against the Code. Deployers may also be asked about their labelling practice. | Record decisions, design and workflow. Template: disclosure statement and internal procedure. |
| 153 | The targeted grandfathering rule covers only marking and detection under Article 50(2). | Retailers acting as deployers get no transition period. |
| 154 | Deep fakes generated or manipulated before 2 August 2026 need no retroactive label. The guidelines encourage labelling without disproportionate effort. | Legacy images: label optional. A new AI edit sets a new date. |
Three reading tips
Read paras. 92 and 116 together
The widely quoted list of standard editing sits at para. 92, in the part on the provider duty under Article 50(2). Whether your image is a deep fake depends on paras. 113 to 116. Para. 116 has its own, similar list of insignificant changes. Read both. “It was only standard editing” is no answer if the image now shows the product differently.
Take the cautious wording seriously
Para. 116 is careful on purpose. A background change is “likely” to have only a minor impact, and whether a change is insignificant depends on the context. Where you have genuine doubts, labelling is the safer choice. Equally, do not label everything by default. An AI label on a real photo is a false statement about where the image came from (see AI labels and conversion).
Examples beat definitions
Three cases from the example boxes are especially clear for stores: an AI product image that shows the product better than in reality (deep fake), a real car against an AI background with no deception (not a deep fake), and a synthetic influencer testing a product (not a work, so the normal label applies). More typical store cases, each with result and citation, are under cases.
What the guidelines leave open
- Old images in new campaigns. For images, para. 154 looks at when they were generated; for text, at when it was published. The guidelines do not say whether reusing an old deep fake in a new campaign triggers the duty. We recommend the label (see deadlines).
- Scenes of use. “Events” includes the depiction of consumer services (para. 113). Whether a scene with your product can mislead remains a case-by-case question.
- Downloading and sharing. The Code allows the label as an overlay on the image. The Commission's icon page, by contrast, summarises that the icon must be visible when content is reshared or downloaded. Para. 115 does not require deployers to plan for sharing by third parties. No authority or court has decided the point.
- Sizes. Neither the Regulation nor the guidelines give pixel values. Design rules are in the Code (see label design).
Using the guidelines day to day
- Classify each AI image with the questions from paras. 12 to 17 (role), 113 to 116 (deep fake) and 154 (timing). The free labelling check asks exactly these questions and cites the paragraph behind every result.
- Write down the result and the citation. That is the evidence para. 148 has in mind.
- Label visibly wherever the image appears, as paras. 142 and 143 require (see AI labels in online stores).
- Review your classification when the formally adopted version is published or a court rules.
Frequently asked questions
Are the Article 50 guidelines legally binding?
No. The guidelines state that they are non-binding. Only the Court of Justice of the EU can interpret the AI Act authoritatively (para. 5). Supervisory authorities, however, follow them. If you read the law differently, you should be able to explain why.
Where can I find the EU AI Act Article 50 guidelines?
On the European Commission's website (digital-strategy.ec.europa.eu), under “Guidelines on the transparency obligations for providers and deployers of AI systems”. The link is in the sources below. Reference: C(2026) 5054 final.
How do the guidelines differ from the Code of Practice?
The guidelines explain how the Commission reads Article 50. The Code of Practice is a voluntary code with concrete measures, such as the “AI” mark and where to place it. Following the Code does not create a presumption that you meet Article 50 (recital 41 of Regulation (EU) 2026/1744). See Code of Practice.
What do the guidelines say about AI product images?
Consumer goods are expressly listed as objects that can be deep faked (para. 113). The guidelines name as a deep fake an AI product image that makes the product look different, more appealing or of higher quality than in reality. A real, unchanged product on an AI background is generally not one, as long as the ad does not mislead about appearance, characteristics or use (para. 116).
Do I have to sign the Code of Practice?
No, signing is voluntary. If you do not sign, you should be able to show by other adequate means how you meet your duties, for example with a gap analysis against the Code (para. 148).
Do the guidelines apply to stores outside the EU?
The AI Act also covers deployers in third countries where the output is used in the Union (Article 2(1)(c)). According to the guidelines it is enough that the deployer foresees distribution in the EU, for example because it sells to EU customers on purpose (para. 13).
Can the paragraph numbers still change?
Yes, that is possible. Formal adoption in all language versions had not been confirmed when we last checked. This page cites the version of 20 July 2026 and will be updated if numbers shift.
Sources
All sources read in full. Paragraph numbers (para.) refer to the European Commission's guidelines on Article 50, C(2026) 5054.
- Commission Guidelines on the transparency obligations under Article 50 AI Act, C(2026) 5054 finalEuropäische Kommission · Commission · 20 Jul 2026
- Regulation (EU) 2024/1689 (AI Act)EUR-Lex · Law · 12 Jul 2024
- Code of Practice on Transparency of AI-Generated Content (final version)Europäische Kommission · Code of practice · 10 Jun 2026
- Commission Opinion on the adequacy of the Code, C(2026) 4839 finalEuropäische Kommission · Commission · 8 Jul 2026
- Regulation (EU) 2026/1744 (Digital Omnibus on AI)EUR-Lex · Law · 24 Jul 2026
- Bundesnetzagentur, press release on the AI ActBundesnetzagentur · Authority · 29 Jul 2026
Not legal advice. This content explains the law based on primary sources; it does not replace an assessment of your individual case.
AI-generated
