The EU AI Act product photography rules are no longer a forecast: since 2 August 2026, Article 50 of the AI Act has applied. These are the transparency rules for content made with artificial intelligence. If you sell online and use AI for product images, this guide explains what you have to label, who is responsible and what to check in your listings, including when your business is based in the UK.
We updated this article in October 2026. It now reflects the European Commission's guidelines on Article 50 (July 2026), the Digital Omnibus on AI (July 2026) and the current rules of Amazon, eBay, Etsy, Google Merchant Center, Allegro, Kaufland and Zalando. We write as a product photography studio, not a law firm: this is not legal advice. If you are unsure about your own case, ask a lawyer.
In short:
- Article 50 of the AI Act applies from 2 August 2026. The Omnibus package moved other deadlines, but not this one.
- Machine-readable marking (metadata, invisible watermarks) is a duty of the AI tool provider, for example the company behind an image generator, not of the shop.
- Whoever publishes a deep fake, an AI image that looks real, has to disclose it visibly. Metadata alone is not enough.
- According to the Commission's guidelines, an AI-generated model and a product made to look better by AI are deep fakes. A real product on an AI background usually is not, as long as it does not mislead.
- UK sellers are covered when their AI images are used in the EU. The UK has no AI labelling law of its own, but misleading product images already break UK consumer and advertising rules.
- Fines for breaching Article 50: up to €15 million or 3% of worldwide annual turnover. For small and medium-sized businesses the lower of the two amounts applies.
What the EU AI Act is and when it applies to product photos
The AI Act is Regulation (EU) 2024/1689 on artificial intelligence. It entered into force on 1 August 2024 and applies in stages. For product images, Article 50 on transparency is the one that matters. Under Article 113 the regulation applies from 2 August 2026, and that date covers Article 50 too (text of the regulation on EUR-Lex).
In July 2026 the EU adopted the Digital Omnibus on AI, Regulation (EU) 2026/1744 of 8 July 2026, in force since 27 July 2026. It moved the rules for high-risk AI systems to December 2027 and August 2028, but it did not change the obligations in Article 50(2) and (4) (consolidated text of 27 July 2026).
Article 50 puts two different duties on two different parties:
- Article 50(2), the provider of the AI tool (for example the company that offers an image generator): outputs must be "marked in a machine-readable format and detectable as artificially generated or manipulated", for instance in metadata or an invisible watermark. This does not apply to tools that only assist with standard editing or do not substantially alter the input data.
- Article 50(4), the deployer, meaning whoever uses AI to create the image (a shop, an agency, a studio): if the image is a deep fake, the deployer must disclose that it has been artificially generated or manipulated.
Each EU country enforces the rules through its own authorities. In Germany, for example, the Bundesnetzagentur supervises the AI Act under the national KI-MIG act (KI-MIG, section 2). In Poland, the rules on inspections and fines of the national AI act apply from 28 October 2026 (Journal of Laws 2026, item 1003).
What needs labelling - and what doesn't
The key term is deep fake. Article 3(60) of the AI Act 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. So it covers product photos, not just faces.
The European Commission explains how to apply this in its guidelines on Article 50 from July 2026. They are not binding, and formally they will apply once they are available in all EU languages (in October 2026 only the English version is published). The table shows the examples that matter most for product photography.
| What you do with the image | Deep fake that needs a label? | Basis |
|---|---|---|
| The whole product image is generated by AI and looks like a photo | usually yes, disclose it visibly | Art. 3(60), Art. 50(4) |
| A model generated by AI wears or holds the product | yes, a realistic AI-generated person counts as a "person" | Commission guidelines, para 113 |
| AI shows the product nicer, in a different colour or of better quality than in real life | yes, and it can also be misleading advertising | guidelines, examples after para 116 and para 129 |
| A real product on a background generated by AI | usually not, as long as it does not mislead about the product | guidelines, examples after para 116 |
| Adding, replacing or removing objects in a way that changes the meaning of the photo | yes | guidelines, examples after para 92 |
| Colour correction, cropping, sharpening, removing sensor dust | no, this is standard editing | Art. 50(2), guidelines, paras 90-92 |
| Removing or blurring a background that is visible in the original, e.g. the white background of a packshot | no, this is standard editing | guidelines, examples after para 92 |
A simple test: could a buyer take the image for a real photo of this product or this person? If yes, and the image was created or substantially changed with AI, label it. Also keep in mind that an "AI" label does not make an image legal if it misleads about how the product looks (guidelines, para 129). Such an image is a problem with or without a label.
Replacing a background with an AI one usually does not turn a real photo into a deep fake. If you would rather not judge such cases at all, use a real photo and classic background removal instead.
Who labels: the tool provider, the shop or the studio
Metadata is the tool provider's job. Companies that offer image generators must mark the outputs in a machine-readable format. Providers whose tools were on the market before 2 August 2026 have until 2 December 2026 to do this (Article 111(4) as amended by the Omnibus). Until then, files from older generators may lawfully have no such marking.
The visible label is the job of whoever uses the AI. According to the guidelines, the deployer is the one who decides whether and how AI is used, for example an advertising agency. A company that only commissions an ad and does not decide on the use of AI is not a deployer. If you choose an AI service yourself, for example a virtual model, you are the one deciding to use it, so ask a lawyer how this applies to you. There is no transition period for deployers: the duty has applied since 2 August 2026.
How to label? The guidelines say plainly that a deployer cannot rely on the metadata added by the generator, because buyers cannot see it (para 117). The information has to be clear and visible at the latest when people first see the image (Article 50(5)). An ad with a realistic AI-generated person presenting a real product does not qualify for the lighter regime for artistic works (paras 122-124).
Penalties
Breaching the obligations in Article 50 can lead to a fine of up to €15,000,000 or up to 3% of the company's total worldwide annual turnover, whichever is higher (Article 99(4)(g)). For smaller businesses the lower amount is the cap:
- SMEs and start-ups: the cap is the lower of the two amounts (Article 99(6)).
- Small mid-caps: since 27 July 2026 the same rule covers them for fines under paragraphs 4 and 5 (Article 99(6a), added by the Omnibus).
- Who fines: national authorities in each EU country, using the amounts set in the AI Act.
These are upper limits, not a price per image. For a shop, platform rules matter more day to day: a listing with a badly labelled image can run into trouble on a marketplace long before any authority looks at it.
UK sellers: what applies to you
The UK is not in the EU, so the AI Act is not UK law. It still reaches UK businesses: under Article 2(1)(c) it applies to providers and deployers in third countries "where the output produced by the AI system is used in the Union". The Commission's guidelines (para 13) explain that this covers deployers who themselves foresee use in the EU, for example by directing distribution there or by posting deep fakes on the globally accessible internet. They are not bound when content reaches the EU through channels they cannot foresee or control.
In practice: if you sell on amazon.de or amazon.fr, list on Zalando or Kaufland, or ship from your own shop to customers in the EU, treat AI deep fakes in your listings the way an EU shop would and label them. This is our reading of the rules, not legal advice.
Northern Ireland: the AI Act does not currently apply there under the Windsor Framework. In August 2026 the UK government wrote that the EU wants it to apply in Northern Ireland, "but the UK has not yet taken a decision" (explanatory memorandum on Regulation 2026/1744). A Northern Ireland seller whose images are used in the EU is still covered by Article 2(1)(c).
UK rules: no AI labelling law, but misleading images are banned
- No duty to label AI as such. As of October 2026 no UK law requires AI images to be labelled. The advertising regulator puts it plainly: "There is no blanket legal requirement in the UK to disclose the use of AI in ads" (CAP News, 29 May 2025). Its test is whether the audience is likely to be misled without a disclosure. It adds that an AI image showing the effect of a cosmetic product that real use would not achieve is likely to be materially misleading, and a disclosure does not fix that.
- CAP Code: under rule 3.1 marketing "must not materially mislead or be likely to do so" (CAP Code, section 3). In March 2026 the ASA ruled against an ad for a robot dog toy whose scenes "appeared to be of real and/or AI-generated puppies" and made the product look highly realistic and lifelike (ASA ruling, 25 March 2026).
- Responsibility: "If an AI tool produces something problematic, the advertiser remains responsible for any output used in their ads" (ASA, June 2026).
- DMCC Act 2024: since April 2025 the CMA can decide itself whether consumer law was broken and fine traders up to 10% of global turnover (CMA, 7 April 2025). A misleading action includes an "overall presentation which is likely to deceive the average consumer" (section 226).
- Consumer Rights Act 2015: goods have to match their description (section 11). A product that falls short of an AI-polished image gives the buyer grounds for a complaint.
So for UK listings the rule is simple: AI must not make the product look different from what the buyer receives. If you sell in both markets, one set of images that works for both is easiest: a real main image, and AI images labelled in the gallery.
How marketplaces handle AI images
Amazon (amazon.co.uk, amazon.de and other stores)
The main image on Amazon is a real photo of the product being sold, with no illustrations, on a pure white background (RGB 255, 255, 255), with the product filling at least 85% of the frame and no extra text, graphics or watermarks (Amazon seller forums, 19 December 2025). A professional studio packshot meets this rule.
In July 2026 Amazon announced that product images and videos containing photorealistic AI-generated people must carry the keyword contains-synthetic-performer in the XMP field dc:subject, added with an IPTC-compatible metadata editor. It applies "in any of our worldwide stores", so also on amazon.co.uk and amazon.de (Amazon announcement). Images without people, or without photorealistic people, are not affected. The trigger was a New York law on synthetic performers in ads (eWeek, 24 July 2026).
eBay (ebay.co.uk, ebay.de)
eBay has no separate rule for AI images. Its picture policy bans "photos that don't accurately represent the item", and under the User Agreement (in force for existing users from 10 August 2026) you are responsible for listing content created with tools such as "generative artificial intelligence tools". An AI image that shows a better item than the one you ship is a listing problem and a returns problem.
Etsy
Etsy's rule is about the item, not the photo: according to its Creativity Standards, sellers must disclose in the listing description if an item is created with the use of AI (as reported by TechCrunch, 9 July 2024; the Etsy page could not be read automatically).
Google Merchant Center
Google requires that all images created with generative AI keep metadata showing this, for example the IPTC tag DigitalSourceType TrainedAlgorithmicMedia, and you must not remove it (Google: AI-generated content). The main image must "clearly display the exact item being sold", and Google rejects overlays and watermarks (image_link specification). Our conclusion: a visible "AI" watermark on the main image can clash with Merchant Center rules. If you mark assets as AI-generated or edited, ads aimed at the EU, India and New York show a visible AI label (AI content label settings).
Kaufland
Kaufland Global Marketplace reads AI information from image metadata on upload and automatically adds a visible AI label. Sellers must keep the metadata unchanged (Kaufland, 23 July 2026). More on images and fees in our guide to selling on Kaufland.
Zalando
Zalando's partner image guidelines (updated 1 October 2026) say: "Don't provide content that requires visible labelling under applicable laws", which includes deep fakes. Zalando may reject AI assets that "fail to represent the product accurately", and AI content needs invisible marking in the metadata by December 2026. The principle behind it: "The product itself is non-negotiable" (Zalando, 25 August 2026).
Allegro
Allegro, the largest marketplace in Poland, has required sellers since 2 August 2026 to tell buyers when an image in a listing is an AI deep fake (Allegro, 13 July 2026). Sellers tick an option when adding the photo and Allegro puts an "AI" watermark on it; the watermark is in beta (Allegro, 29 July 2026). Whether an image is a deep fake is for the seller to judge, and the seller is responsible for what the listing shows (Allegro help).
| Platform | AI images | Main image | Since |
|---|---|---|---|
| Amazon | photorealistic AI people: keyword contains-synthetic-performer in metadata | real product, white background, 85% of the frame | July 2026 |
| eBay | no AI rule; you are responsible for AI-made content | must accurately represent the item | User Agreement from 10.08.2026 |
| Google Merchant Center | keep IPTC TrainedAlgorithmicMedia; AI label in ads in the EU | exact item being sold, no watermarks | help page from February 2024 |
| Kaufland | automatic visible AI label from metadata | product on a white background | 23.07.2026 |
| Zalando | no content that needs a visible label; invisible marking by December 2026 | product shown accurately | 01.10.2026 |
| Allegro | declaration for deep fakes and "AI" watermark (beta) | real product | 02.08.2026 |
IPTC, C2PA and metadata: how to mark AI images technically
Metadata does not replace a visible label on a deep fake, but you need it for two reasons: Google Merchant Center requires it, and files from generators increasingly contain it and must not be stripped. The AI Act itself lists watermarks, metadata and cryptographic methods among the possible techniques (recital 133).
IPTC Digital Source Type
The IPTC Digital Source Type vocabulary describes where an image comes from. The values you will meet with product photos:
digitalCapture- a photo taken with a camera of a real scenecomputationalCapture- several frames merged automatically without generative AI, e.g. HDR on a phonealgorithmicallyEnhanced- an algorithmic fix that does not change the content, e.g. sharpening or noise reductioncompositeWithTrainedAlgorithmicMedia- an image edited with generative AI, e.g. inpainting or outpaintingcompositeSynthetic- a composite in which at least one element comes from generative AItrainedAlgorithmicMedia- an image fully generated by AIalgorithmicMedia- an image made purely by an algorithm without training data, e.g. a mathematical pattern (this is not HDR, upscaling or a 3D render)
Google Merchant Center says that images created with generative AI must carry TrainedAlgorithmicMedia, and of the composite values it only names CompositeSynthetic. For Merchant Center it is safer to use one of these two. You can add the tag in Photoshop, Lightroom or the free ExifTool, for example: exiftool -XMP-iptcExt:DigitalSourceType="http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia" image.jpg
C2PA (Content Credentials)
C2PA is a standard for a file's "passport": it records who created or changed an image and with which tool, and it is cryptographically signed. Adobe announced that Firefly-powered features in Photoshop automatically attach Content Credentials to the file (Adobe, 23 April 2024). Check what was actually written into your own file before you build a process on it.
Watch out when publishing: image optimisation plugins, CDNs and some platforms strip metadata during compression, and so do screenshots and format changes (e.g. PNG to JPEG). After uploading, download the image from the page and check with ExifTool whether the tags survived.
Studio photos: when Article 50 does not apply
A photo taken with a camera and edited in the usual way (colour, crop, retouching without generating new elements) is not AI-generated content, so it does not need an Article 50 label. The line runs at generative tools: if AI adds or removes elements during retouching in a way that changes what the photo shows, the assessment can be different.
At marszalstudio we photograph real products in our studio in Warsaw. In the app you can add the "Statement: no generative AI" option to your order (€7 per shoot): a signed studio statement that no frame or part of a frame comes from a generative model. We also offer AI services and say so openly: your garment on an AI-generated model, made from our apparel packshot, and AI retouching. They are separate options and cannot be combined with the statement. An image with an AI model is a deep fake of a person according to the Commission's guidelines, so it has to be labelled visibly when you publish it. Prices are in our product photography pricing and in the app.
| Studio photo, classic editing | Image from generative AI | |
|---|---|---|
| Article 50 label | not needed | visible disclosure when the image is a deep fake |
| IPTC metadata | digitalCapture | TrainedAlgorithmicMedia or CompositeSynthetic (Google) |
| Amazon main image | meets the "actual product" rule | not as the main image |
| Allegro | no watermark | declaration and "AI" watermark if it is a deep fake |
What works best is two files per product: a real photo on a white background as the main image and, if you want one, an AI image in the gallery, labelled according to the platform's rules. With food photography, jewellery photos or clothing, product accuracy is exactly what the buyer pays for. Techniques such as focus stacking combine real exposures and need no label.
Checklist: what to do now
Whether or not you use AI for product images, go through this list for your listings:
- List the images that were created or substantially changed with AI, including those from suppliers and agencies.
- Check each image with the test from this guide: could a buyer take it for a real photo of the product or a person? AI models, products generated from scratch and products made to look better by AI need a label.
- Label visibly: with a caption or badge next to the image in your shop, or with the platform's own tool (e.g. on Allegro), visible from the first moment people see the image.
- Do not strip AI metadata from files, and check that your shop system, plugins and CDN keep it (a Google Merchant Center requirement).
- Keep the main image real: authentic packshots on a white background meet the rules of Amazon and Google.
- Agree in writing with your agency or studio who decides on the use of AI and who labels the images.
- Follow the changes: the formal adoption of the Commission's guidelines in all languages, the voluntary code of practice on marking and labelling AI-generated content (June 2026) and the optional EU icons for AI content.
Frequently Asked Questions
Does the EU AI Act apply to UK businesses?
Yes, when your AI images are used in the EU. Article 2(1)(c) covers deployers outside the EU where the output is used in the Union, and the Commission's guidelines (para 13) include posting deep fakes on the globally accessible internet. If you sell to EU customers, label AI deep fakes the way an EU shop would. In the UK itself there is no AI labelling law, but misleading images break the CAP Code and the DMCC Act 2024.
How do I know if AI-generated product images will be accepted by Amazon, eBay and Google Shopping?
Check three sets of rules. Amazon: the main image must be a real photo of the actual product on white, and photorealistic AI-generated people must be tagged with contains-synthetic-performer in the image metadata. eBay: there is no AI ban, but the photo must accurately represent the item, and you are responsible for content made with generative AI tools. Google Merchant Center: AI images must keep the IPTC TrainedAlgorithmicMedia tag, and the main image must show the exact item being sold. The set-up that works on all three is a real photo as the main image and AI images only in the gallery.
Does removing a background count as AI-generated?
No, if you remove or blur a background that is visible in the original photo: the Commission's guidelines treat this as standard editing. Replacing it with an AI background usually does not make a deep fake either, as long as it does not mislead about the product. If the tool wrote AI metadata into the file, keep it: Google and Kaufland read it.
Does an AI-generated model need a label?
Yes. According to the Commission's guidelines, a realistic AI-generated person counts as a "person" in the deep fake definition, even if it does not resemble anyone real. An ad with such a person presenting a real product does not qualify for the lighter regime for artistic works. On Amazon you also need the contains-synthetic-performer keyword in the file.
Do product images published before 2 August 2026 need a label?
Not under the AI Act. The Commission's guidelines (para 154) say that deep fakes generated or manipulated before 2 August 2026 do not need to be labelled retroactively, although the Commission encourages it where it takes no disproportionate effort. AI images created from 2 August 2026 need a label. Platform rules can be stricter, so check them for your live listings.
What about 3D renders and CGI product images?
A classic 3D render made in Blender or KeyShot without generative AI is not AI-generated content, so Article 50 does not apply to it, but like any image it must not mislead about the product. The IPTC value algorithmicMedia means media made purely by an algorithm without training data, such as a mathematical pattern, so it is not a label for renders. If generative AI adds details or a scene to a render, judge the result like any other AI image.
Does focus stacking or HDR require AI labelling?
No. Focus stacking and HDR merge several real exposures without generating new content. In the IPTC vocabulary, HDR on a phone is computationalCapture, which is still a photo of a real scene.
Summary
Article 50 of the EU AI Act has applied since 2 August 2026. A shop does not have to add metadata to every photo itself, but it has to label AI images that look real, visibly. Marketplaces add their own rules on top: Amazon, Google, Kaufland and Allegro already have them.
- Go through the images in your listings and mark those that were created or substantially changed with AI.
- Label AI models, products generated from scratch and products made to look better by AI visibly.
- Treat a real product on an AI background with care: usually it is not a deep fake, as long as it does not mislead.
- Do not strip AI metadata from files, and keep a real photo as the main image.
- UK sellers: the AI Act applies when your images are used in the EU, and UK consumer law already bans misleading images.
- Fines: up to €15 million or 3% of turnover, for SMEs the lower amount.
If you are looking for a studio that delivers real photos with a clear origin, take a look at marszalstudio. We photograph products of every size: packshots on a white background, jewellery, cosmetics, clothing and food. In the app you can add the "Statement: no generative AI" to your shoot (€7 per shoot). Questions? Contact marszalstudio.
Sources
- Regulation (EU) 2024/1689 (AI Act): Art. 2(1)(c), Art. 3(60), Art. 50, Art. 99, Art. 113, recitals 133-134; consolidated text of 27 July 2026
- Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI)
- European Commission: guidelines on the transparency obligations under Article 50, July 2026 (paras 13, 14, 90-92, 113, 116, 117, 122-124, 129, 154)
- European Commission: code of practice on marking and labelling AI-generated content, 10 June 2026; EU icons for AI-generated content
- Germany: KI-MIG, section 2; Poland: act on artificial intelligence systems, Journal of Laws 2026, item 1003
- UK: CAP News, disclosure of AI in advertising, 29 May 2025; CAP Code, section 3; ASA ruling on UAB CommerceCore, 25 March 2026; ASA, AI and deepfakes, June 2026
- UK: Digital Markets, Competition and Consumers Act 2024, section 226; CMA, new consumer protection regime, 7 April 2025; Consumer Rights Act 2015, section 11; UK explanatory memorandum on Regulation 2026/1744 (Northern Ireland)
- Amazon seller forums: product image requirements, 19 December 2025; Amazon: tagging AI-generated people, July 2026; secondary source: eWeek, 24 July 2026
- eBay UK: picture policy; eBay UK: User Agreement; Etsy: Creativity Standards (via TechCrunch, 9 July 2024)
- Google Merchant Center: AI-generated content; image_link specification; AI content label settings
- Kaufland Global Marketplace, 23 July 2026; Zalando partner image guidelines, 1 October 2026; Zalando, 25 August 2026
- Allegro help: labelling AI images; Allegro for developers, 13 July 2026; 29 July 2026
- IPTC NewsCodes: Digital Source Type; Adobe, 23 April 2024 (Content Credentials in Photoshop)
