EU AI Act AI-generated product images · 18 min read
EU AI Act Article 50 and AI Fashion Product Images: A Buyer's Guide to Labels and Marking
Article 50 of the EU AI Act, from 2 August 2026, splits fashion-image duties: the AI provider marks the file, and the deployer — usually the brand — shows a visible label when an image is a deepfake. This page triages common jobs from the Commission's own examples.
By Surajnarayanan SUpdated
Not legal advice. This guide summarises the EU AI Act and the European Commission's published Article 50 guidance as read on 5 October 2026. The guidance is non-binding, and how it applies to a specific image depends on context. Confirm your labelling decisions with your own legal counsel.
At a glance
- Provider marks the file (machine-readable); deployer shows the label (human-visible) — one does not replace the other
- Triage per image: standard edit, AI background, AI model, face replacement, misrepresented product
- Merchant Center IPTC tags are a separate feed rule, not an EU visible label
8M Studio is a managed AI fashion production studio, not a law firm. This page is informational and is useful even if you never work with us.
Direct answer
From 2 August 2026, EU AI Act Article 50 splits the job. The AI system provider must mark generated or manipulated images in a machine-readable way. The deployer — usually the brand that chose AI and publishes the image — must visibly disclose deepfakes at first exposure. Realistic AI on-model images may meet that definition, depending on context. Standard edits such as colour correction usually don't. A managed studio can handle marking-ready files and QC; the brand owns the visible label.
Best for
- +Fashion ecommerce, brand and creative-ops leads publishing AI on-model, AI-edited or AI campaign images to EU shoppers
- +UK and US brands whose storefronts, marketplaces or ads reach the EU
- +Buyers who need vendor questions and a per-image triage before briefing counsel
Not designed for
- -Legal advice or a compliance certificate
- -A full explainer of every AI Act obligation (chatbots, emotion recognition, high-risk systems)
- -Google Shopping metadata detail — that is the Merchant Center AI product images guide
- -Plugin or tool recommendations
Article 50 creates two separate duties
Article 50 of the EU AI Act applies from 2 August 2026. For fashion imagery it creates two separate duties. The Commission's Article 50 guidelines (July 2026) are published guidance and are non-binding (point 5). This page quotes them. It does not treat them as a ruling.
The provider marks. Providers of AI systems that generate synthetic image content “shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated” (Article 50(2)).
The deployer discloses. Deployers of an AI system that generates or manipulates image content “constituting a deep fake, shall disclose that the content has been artificially generated or manipulated” (Article 50(4)). That disclosure must be given “in a clear and distinguishable manner at the latest at the time of the first interaction or exposure” (Article 50(5)).
The Commission is explicit that the two don't substitute for each other: deployers “cannot simply rely on the machine-readable marking embedded in the content by the provider under Article 50(2) of the AI Act to fulfil their disclosure obligation.”
Not every AI-touched image is a deepfake, and not every edit needs a mark. The rest of this guide sorts fashion images into those buckets, quoting the Commission's own examples.
Who is the provider and who is the deployer in a fashion image workflow
Provider
The Commission defines providers as those that “develop AI systems, or have them developed, and place them on the EU market or put them into service under their own name or trademark,” including providers outside the EU “if the output of their AI system is used in the EU.” In practice that is the company offering the image-generation system: a model or tool vendor, or a studio that builds and runs its own system under its own name. The guidelines add that providers may rely on an upstream model provider's or a third party's marking solution, “without prejudice to the responsibility of the provider of the AI system to demonstrate compliance” (point 74).
Deployer
Deployers are those “using AI systems under their authority,” and the guidelines read authority as “responsibility over the decision to deploy the system and over the manner of the actual use of the system (including its outputs)” — technical control is not required (point 12). Employees acting for a company are not separate deployers, and “a legal person remains a deployer even if it involves third parties (e.g. contractors, freelancers)” acting on its behalf and under its control (point 14).
Where the brand usually sits
When a brand decides to use AI for its on-model catalogue, approves the outputs and publishes them on its storefront, it is very likely the deployer for those images. The guidelines give one contrast case: “a company that merely commissions an advertising agency to produce an advertisement, without taking decisions and exercising control over whether and how the advertising agency uses AI in the production process, is not a deployer” (point 14). Most AI catalogue programmes are the opposite of that — the brand chose AI. Write the roles into the contract rather than assuming.
Where platforms sit
Hosting services and online platforms whose role is limited to disseminating content created by others “are not deployers” (point 16). Your marketplace or storefront platform will not carry your label duty for you. The guidelines suggest deployers in complex distribution chains take “proportionate measures” so the label still displays at first exposure, for example “via contractual conditions with distributing partners, user experience (UX) settings and interfaces” (point 12).
Outside the EU
UK and US brands are in scope as deployers where they foresee use of the output in the EU, “including by posting deep fakes on the globally accessible internet” (point 13). Content reaching the EU through channels that are “unforeseeable and outside their control” is not covered.
What the AI provider or vendor must embed: machine-readable marking
Article 50(2) is a provider duty. The mark must be machine-readable and detectable, and the technical solution “effective, interoperable, robust and reliable as far as this is technically feasible.” The guidelines list example techniques from Recital 133: “watermarks, metadata identifications, cryptographic methods for proving provenance and authenticity of content, logging methods, fingerprints or other techniques, and a combination of such techniques” (point 73). Providers “are not required to record or keep a full provenance chain” (point 73).
Timing for existing tools
For AI systems placed on the market before 2 August 2026, the Commission says providers “must comply with those obligations only as from 2 December 2026.” That grace period applies only to marking and detection under Article 50(2). It does not delay the deployer's visible-label duty, which applies from 2 August 2026.
Code of Practice
The Commission's Code of Practice on Transparency of AI-generated Content is a voluntary tool to demonstrate compliance with marking and labelling. The Commission FAQ says it “has been assessed as adequate by both the Commission and the AI Board.” Providers that don't sign “will have to demonstrate compliance through alternative adequate means.”
What this means for a buyer
Marking is the vendor's obligation, but you inherit the problem if files arrive unmarked or lose the mark before publication. Ask in writing (see the vendor questions below).
The AI image label checker shows whether files carry IPTC or C2PA marking; it can't decide whether a visible label is needed.
Where a managed studio fits
A managed provider can deliver files with machine-readable AI-origin metadata already written, keep that metadata through its own export path, and keep a per-image record of which frames were generated, which were edited, and how. Whether any specific marking method satisfies Article 50(2) is assessed against the provider's obligations and the Code of Practice — ask your vendor which method they use and on what basis.
What the brand must show: the visible label
Where an image is a deepfake under the Act, the deployer must disclose it.
What the label must do
The Commission FAQ: disclosure should happen “in a clear and distinguishable manner. It should be understandable and perceivable by natural persons (e.g. with visible or audible labels), without need for any specific technical tools or performing dedicated actions.” The guidelines add that information is not clear and distinguishable “where it can be easily overlooked or missed… (e.g. only included as part of a manual or hidden under layers of menu options on an online interface, part of terms of use that are often not read by users)” (point 142).
When
“At the latest at the time of the first interaction or exposure” (Article 50(5)). For content, the guidelines say the obligation “applies to each output of an AI system with respect to any natural person exposed to the content” (point 143). On a PDP, that points to the label living on or beside the specific image, not on a policy page.
Per image, not per site
Because the duty attaches to each output, a gallery with one real photo and three AI on-model frames needs the disclosure on those three. A footer statement or a site-wide notice is, at best, supporting information.
The EU icons
The Commission publishes a free set of icons with three types: a Basic icon, Fully AI-Generated (“no human-created elements or human editorial control (apart from prompting)”), and Partially AI-Modified (“pre-existing, human-made content was partially modified with AI”). Their use “is optional, but the labelling requirements under Article 50 AI Act are not,” and “the use of these icons does not establish legal compliance by itself.” User testing found “performance improved across all measures when the basic icon was accompanied by a text label (e.g. modified).”
Placement summary from the Commission's icon page
- Clearly perceivable “at the latest at the time of first exposure”
- Placed “where no intervening overlay elements exist”
- “Directly embedded into the deepfake… unless equivalent alternatives are available such as a user interface overlay”, and visible “when content is reshared or downloaded”
- Accessibility: clearly visible size; plain language avoiding abbreviations other than “AI”; alt text or ARIA labels where possible
- Signatories of the Code commit to its detailed placement specifications. Non-signatories still have to meet Article 50(5).
Practical note on language
The icon page asks for plain language without jargon. For multi-market EU storefronts, localising the text label per market is a sensible default to agree with counsel. It is not a quoted rule on this page.
When an AI on-model image may count as a deepfake (and when it may not)
This is the question fashion buyers care about most, and the one most easily overclaimed. Here is what the text says.
The definition
Article 3(60): a deep fake is “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 Commission's cumulative criteria
The Commission FAQ sets three cumulative criteria: resemblance (“a high level of similarity”), existing, and “false appearance to be authentic or truthful.”
Existing is broader than a real person
The guidelines say “it is sufficient for simulated persons… to resemble someone or something that exists, can plausibly exist or could have plausibly existed in reality” (point 113). “Persons” includes “realistic AI-generated human avatars or personas”. “Objects” includes “consumer goods”. So a photoreal synthetic model wearing a real garment can meet the first criteria without resembling any real individual.
False appearance depends on context
It is assessed “as a whole, taking into account the level of resemblance, the potential substantive message of the content, the intended and foreseeable deployment contexts… and the intended and reasonably foreseeable audience composition and their expectations”. It “does not require the intention of the deployer to deceive” (point 114). The guidelines say the use of AI for “essential elements that impact audience perception is likely to” create a false appearance, “such as fully AI-generated actors.” Also: “a high degree of photorealism renders it more likely that such content should be considered a deep fake… but photorealism alone is not determinative” (point 114).
Commercial images are not automatically creative works
The lighter disclosure regime for “evidently artistic, creative, satirical, fictional or analogous” works is read strictly. It excludes content whose nature “is exclusively informative or commercial and is recognisable as such”. Advertisements “might be regarded as evidently creative or fictional works in certain, specific situations, but not in others” (point 122). Among the guidelines' examples that are not creative works: “a realistic synthetic influencer testing out a sponsored real product, focused solely on displaying the product functionalities.”
What that means for a PDP
A photoreal AI-generated model presenting a real garment on a product page is a realistic person in a commercial setting where shoppers generally expect to see the product as it is. On the Commission's reading, that kind of image may well meet the deepfake definition. Many brands will reasonably treat photoreal AI on-model PDP images as needing a visible label by default, unless counsel concludes otherwise for a specific context. This is a cautious default, not a ruling. The text leaves the final assessment to the deployer, case by case.
Where the guidelines point the other way
“Simulated persons, objects, places… that, for example, defy the laws of nature or physics… and have no potential to mislead are considered unrealistic.” Clearly stylised, illustrative or surreal campaign art is weaker deepfake territory. Clearly creative campaign work may fall under the lighter “appropriate manner that does not hamper the display or enjoyment of the work” regime. Both are case-by-case calls.
What likely does not need a label: standard retouching, colour and background work
Two separate tests apply here. Don't merge them.
Test 1 — does the provider's marking duty apply?
Article 50(2) “shall not apply to the extent the AI systems perform an assistive function for standard editing or do not substantially alter the input data provided by the deployer or the semantics thereof.” Editing goes beyond standard editing “if the content is changed in a material way… that affect[s] its meaning, style or intent” (point 90).
The Commission's examples benefitting from the exception include (point 92): “Minor cropping, minor colour adjustments or corrections, lightening or darkening, sharpening for enhanced clarity”; “Removal of dust spots caused by a dirty lens or sensor… deleting and obscuring backgrounds that are visible in the original file”; “Standard adjustments to colour and contrast… minor corrections to level the horizon”; and “Pixel filling for aspect ratio adaptation or enhanced clarity… other non-substantive edits for cosmetic and technical purposes.”
Examples that require marking include: “Removal, replacement or insertion of objects or persons in existing images… that changes meaning and substance of the content; face replacement or substantial facial modification”; “altering the body shape or the skin colour of a person; extreme lightening, darkening, colour and contrast changes that change the meaning, intent and messaging”; and “Creation of composite images… that modifies the representation of persons, objects, events or facts.”
Test 2 — is it a deepfake the deployer must label?
The guidelines say AI manipulation of “insignificant substantive or technical aspects of pre-existing content” may not create a deepfake. For example: “AI-powered colour correction, background extensions of existing content, adjustments or replacements of backgrounds for clearly aesthetic purposes, compositions and arrangements of existing products, or re-scaling of images applied in product advertisements or packaging is likely to have only a minor impact on a person's perception of the authenticity and truthfulness of the advertisement and the product” (point 116). Among the non-deepfake examples: “A real product (e.g., a car) shown in an advertisement against an AI-generated background… as long as the ad is not likely to mislead the audience about the product's actual representation.”
In fashion terms
Routine retouching on a real photo, such as dust removal, colour correction to match the garment, minor crops, a cleaned or knocked-out background, or aspect-ratio fill for channel crops, is likely outside the visible-label duty. The milder edits are also likely outside the marking duty. A real garment photographed on a real person against an AI-replaced background “for clearly aesthetic purposes” is likely not a deepfake, though the marking question may still apply to the generated background. Changing the person (body shape, skin colour, face) or the product is a different category.
Triage table: common AI fashion image jobs
Informational. Each row is a starting point for your counsel, not a determination. “Likely” stays in every cell. Flat-lay and mannequin inputs that become photoreal on-model frames are the route in mannequin and flat-lay to on-model at scale. A colourway swap is the question in AI colorway swaps for an on-model catalogue: whether the colour shown is true to the real SKU. A photoreal campaign family is the job in AI campaign image families. Holding one synthetic person across a book is fashion AI identity lock.
| Image job | Provider marking (Art. 50(2)) — likely | Visible deepfake label (Art. 50(4)) — likely | Why (Commission text) |
|---|---|---|---|
| Real photo: colour correction, dust removal, minor crop, sharpening | Likely not required (standard editing) | Likely not required | Point 92 standard-editing list; point 116 |
| Real photo: background removed or obscured, aspect-ratio pixel fill | Likely not required for marking | Likely not required for the visible label | Point 92 (“deleting and obscuring backgrounds…”, “pixel filling for aspect ratio”) |
| Real garment and real model, AI-generated or replaced background for aesthetics | Possibly required for the generated content; ask the vendor | Likely not required if the product is not misrepresented | Point 116; car-background example |
| Mannequin, flat-lay or ghost to AI-generated photoreal on-model | Required (generated content) | Likely required by default; context-dependent | Points 113–114 (“plausibly exist”, “fully AI-generated actors”, photorealism) |
| Face or identity replacement on an existing model photo | Required (“face replacement”) | Likely required | Point 92; EU icon “Partially AI-Modified” face-swap example |
| Body shape or skin colour altered by AI | Required for this alteration | Likely required for this alteration | Point 92 |
| AI colourway swap on an approved on-model frame | Depends on how substantially the input is altered; ask the vendor | Depends whether the colour shown is true to the real SKU | Points 91–92; point 116 product-ad caution |
| Photoreal AI campaign family with synthetic talent | Required for generated content | Likely required unless evidently creative; then a lighter “appropriate” disclosure | Points 119–123 |
| Any image that makes the product look different or better than it is | Do not treat marking as a way to ship it | Listed as a deepfake example. Do not label and ship: hold it. | Point 116 examples |
| Images generated before 2 August 2026 | Not retroactive for marking | Not retroactive (relabelling encouraged where possible) | Commission FAQ; point 154 |
How this fits with the Google Merchant Center AI image rule
They are different regimes doing different jobs. You need both.
Merchant Center says generative-AI product images must carry IPTC DigitalSourceType metadata identifying AI origin, and that metadata must not be removed. It is machine-readable and lives in the file Google crawls. The full rule, and why CDNs strip it, is in our Google Merchant Center AI product images guide.
The EU visible label is for people. The Commission is explicit that metadata does not satisfy it: deployers “cannot simply rely on the machine-readable marking.” A file with a perfect IPTC tag and no visible label does not meet the deployer duty on an image that is a deepfake.
Where they can collide: Merchant Center's product-image rules restrict promotional overlays on product images. Baking a visible “AI-generated” badge into the image_link pixels risks breaking feed rules. The EU icon page allows “equivalent alternatives… such as a user interface overlay”. A common pattern to evaluate with counsel and feed ops is below. Follow current Merchant Center Help for overlays.
- Storefront PDP: a UI-overlay label on each AI image, plus alt text or ARIA.
- Shopping feed file: a clean image with IPTC DigitalSourceType preserved. Follow current Merchant Center Help for overlays.
- Social, ads and marketplaces: a label that survives the channel, set via your agreements and UX settings with that channel (point 12). Some marketplaces won't take a visible label at all: Zalando bans on-image AI labels, as explained in Zalando's AI-generated product image rules.
One metadata pipeline, two outputs
The production record that tells you which IPTC value to write is the same record that tells you which images need a visible label. Keep it per image.
Amazon adds a third, separate signal: a contains-synthetic-performer keyword for photorealistic AI-generated people. It is a metadata keyword, so on its own it is not a visible label either. See Amazon's AI-generated model image tag.
A label does not fix a wrong garment
The guidelines list as a deepfake example “an AI-generated image of a product in advertisement or packaging that can affect the audience's perception and mislead as to the actual product appearance, characteristics or use (e.g. making the product appear not identical to the real product, more appealing or with improved quality than in real life).” Article 50 is also “without prejudice to other transparency obligations laid down in Union or national law” (Article 50(6)), and the guidelines treat the deepfake test separately from deception under the Unfair Commercial Practices Directive.
For fashion, that is the same failure as print drift, a wrong colour, collapsed lace or a changed silhouette. The answer is not a bigger label. Hold the frame. The pattern is why DIY AI fashion images fail on garment fidelity. Use the AI fashion imagery QC checklist: Gate 2 (product and garment fidelity), Gate 5 (channel and technical compliance) and Gate 6 (delivery and audit trail). When AI is the wrong tool for the garment, start with when not to use AI fashion images.
Vendor questions to put in writing
Ask every AI image provider or studio before the next EU-bound delivery.
- Which AI systems produce our images, and who is the provider of each for Article 50(2)?
- What machine-readable marking is applied (for example metadata, watermark or a provenance manifest), at what stage, and on what basis do you consider it compliant? Are you relying on an upstream model provider's marking?
- Do delivered files keep that marking through your export? Which of our steps (DAM, CDN, theme optimiser) could strip it?
- Do you write IPTC DigitalSourceType for Merchant Center, and which value per image?
- Will you deliver a per-image record: generated, AI-modified, or standard-edited only, with the method?
- Which images involve face replacement, body or skin alteration, or a synthetic person? Flag them for our labelling decision.
- Who decides and applies the visible label on each surface: us, you, or a channel partner? Put it in the statement of work.
- What happens to frames that misrepresent the product? Confirm they are held, not delivered with a label.
Where 8M Studio fits
8M Studio is a managed AI fashion production studio. It is not a self-serve generator, not a labelling plugin, and not a law firm.
What a managed provider can own, within an agreed scope: delivery files with AI-origin metadata already written (including IPTC DigitalSourceType for Shopping where in scope), metadata kept through the studio's own export, a per-image production record (generated, AI-modified or standard edit), and human QC that holds frames which misrepresent the garment.
What the brand still owns: the decision that a given image needs a visible label, the label's wording and placement on its storefront and channels, and the legal read with its counsel. There is no rate card on this page. Channel visual specs for Shopify, Amazon Fashion and Zalando-class catalogues are a separate matrix: fashion ecommerce image guidelines by channel. This page stays on transparency.
UK brands selling into the EU: how we work with UK fashion brands.
- AI Catalogue Production — recurring on-model and PDP volume for EU storefronts, delivered with marking-ready files and QC.
- AI Campaign Production — art-directed image families. Ask early whether the family is photoreal synthetic talent; it changes the label decision.
- Custom AI Production Systems — when labelling metadata, production logs and routing rules must live inside your own governed pipeline.
Buyer checklist for EU-bound AI fashion images
- Inventory live and upcoming images per SKU: real photo, standard-edited, AI-modified, fully AI-generated.
- Note the generation or manipulation date. Images made before 2 August 2026 are not retroactive.
- Triage each AI image with the table above, and get counsel's sign-off on your default for photoreal AI on-model images.
- Confirm in writing who the provider is and what marking they apply. Ask about the 2 December 2026 grace only for systems on the market before 2 August 2026.
- Decide label wording, and whether to use the EU icons with a text label, per market.
- Place the label per image at first exposure on each surface: PDP overlay, alt text or ARIA, social, marketplace.
- Keep IPTC DigitalSourceType on Shopping files, and keep visible badges out of image_link pixels unless current Merchant Center Help allows them.
- Hold, don't label, any frame that misrepresents the product.
- Keep a per-image record of who decided, what was applied, and when.
Not legal advice. Confirm decisions with your counsel.
Frequently asked questions
Do AI-generated fashion product images need a label under the EU AI Act?
Only images that meet the Act's deepfake definition need a visible label from the deployer, but photoreal AI on-model images may well meet it. The Commission's guidelines say a simulated person who “can plausibly exist” can satisfy the definition, and that photorealism makes it more likely, though not automatically. Many brands treat photoreal AI on-model PDP images as needing a label by default, with counsel's sign-off.
When does Article 50 of the EU AI Act apply?
Article 50 applies from 2 August 2026. The Commission says providers of AI systems placed on the market before that date have until 2 December 2026 for the machine-readable marking duty only. The deployer's visible-label duty is not delayed.
Who is responsible: the AI vendor, the platform or the brand?
The provider of the AI system must mark outputs in a machine-readable way. The deployer, usually the brand that decided to use AI and publishes the image, must visibly disclose deepfakes. The Commission says platforms that only disseminate others' content are not deployers; write your vendor's role into the contract.
Is embedded metadata or a watermark enough?
Not for the deployer's duty. The Commission says deployers “cannot simply rely on the machine-readable marking” and that labels should be perceivable without technical tools. Metadata serves the provider's marking duty and channel rules such as Merchant Center.
Does colour correction or background removal need a label?
Probably not. The Commission lists minor colour corrections, minor cropping, dust removal and “deleting and obscuring backgrounds that are visible in the original file” as standard editing. It also says aesthetic background replacements in product ads are likely to have only a minor impact on perceived authenticity, unless the product is misrepresented.
Does a face swap or AI model replacement need a label?
Likely yes. The Commission lists “face replacement or substantial facial modification” and altering body shape or skin colour as changes that require marking. Its “Partially AI-Modified” icon example is a face swapped in an authentic photograph.
How does this relate to Google Merchant Center's AI image rule?
They are separate. Merchant Center requires IPTC DigitalSourceType metadata on generative AI product images; the EU rule requires a human-visible label on deepfakes. Use a storefront overlay for the label and keep feed images clean with metadata intact, checking current Merchant Center Help.
Do we have to relabel images made before 2 August 2026?
No. The Commission says content generated before 2 August 2026 does not need to be labelled retroactively. It encourages deployers to do so where possible, without disproportionate effort.
Can 8M Studio make our images EU AI Act compliant?
No studio can make that call for you. 8M Studio, as a managed production studio, can deliver marking-ready files, a per-image production record and human QC within an agreed scope. The brand still owns the visible label and the legal assessment.
Evidence and sources
- EUR-Lex — Regulation (EU) 2024/1689 - Article 3(60) deep fake definition, Article 50, and Article 113 application from 2 August 2026. Reviewed against Commission guidance as of 2026-10-05. Verified 2026-10-05.
- European Commission FAQ — transparency obligations under Article 50 - Provider and deployer duties, the deepfake criteria, the standard-editing exception, 2 August 2026 applicability, and the 2 December 2026 grace for marking only. Verified 2026-10-05.
- European Commission — guidelines on transparency of AI-generated content - Policy page for the Article 50 guidelines, C(2026) 5054, July 2026. Non-binding (point 5). This page does not treat the guidelines as formally adopted legislation. Verified 2026-10-05.
- European Commission — EU icons for labelling AI-generated content - Optional Basic, Fully AI-Generated and Partially AI-Modified icons. Use does not establish legal compliance by itself. Verified 2026-10-05.
Continue the decision
Triage the next EU-bound drop
Not legal advice. If your next EU-bound drop includes AI on-model or AI-edited images, triage them per image, agree the visible label with counsel, and get your vendor's marking in writing. Book a production-fit call when you want AI Catalogue Production delivered with marking-ready files, a per-image record and human QC.