Creating AI images for store clients: what agencies, photographers and freelancers need to know about labelling

Research as of 1 October 2026. This guide is for Shopify agencies, product photographers, content studios and freelancers who create or edit images with AI for retailers. It explains who owes the labelling under the AI Act, what you should hand over to your client and which points belong in the contract. The retailer’s view of the same question is in the guide AI stock images in your online store, the basics are in the guide Do I have to label AI images in my online store?. For the wider picture, see the labelling rules in Article 50. General information, not legal advice for your individual case.

Who is the deployer: you or your client?

Article 50(4) of the AI Act (Regulation (EU) 2024/1689) requires the visible labelling from “deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake”. Under Article 3(4), a deployer is a person or body “using an AI system under its authority”, except where the system is used in the course of a personal non-professional activity. What matters, therefore, is who uses the AI system, not who publishes the finished image. The Digital Omnibus (Regulation (EU) 2026/1744) changes none of this: Article 3 is amended only in point 14 and supplemented by points 14a and 14b (Article 1 point 4), Article 50 only in paragraph 7 (Article 1 point 20).

What “under its authority” means is explained in the European Commission’s guidelines on Article 50 (annex to C(2026) 5054 final of 20 July 2026):

• Para. 12: it means assuming responsibility “over the decision to deploy the system and over the manner of the actual use of the system (including its outputs)”. Technical control over the system is not required, as long as the deployer decides for what purposes and how the system is used.
• Para. 14, first sentence: where the deployer is a legal person, such as “an advertising company”, the employees who act under its instructions and control are not separate deployers. The guidelines give as examples “digital animators, web designers, content creators, journalists”.
• Para. 14, second sentence: “A legal person remains a deployer even if it involves third parties (e.g. contractors, freelancers) in the operation of the system on its behalf and under its responsibility and control.”
• Para. 14, counter-example: “By contrast, 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.”

This gives the typical constellations. Where the guidelines do not name a case expressly, the right-hand column says so:

Many jobs fall in between: the client wants “something with AI” and leaves the execution to you, or specifies a look and picks from your variants. The guidelines do not say where deciding on the “how” begins. Nor do they say whether, in the first row of the table, you are also a deployer yourself; they assign the obligation to the client there. In practice: agree with your client before the job who decides on the use of AI, and put it in writing. Briefings, approvals and the choice between variants will later show how things actually went.

Freelancers. The exception for personal non-professional use does not help you. Under para. 19 of the guidelines, any activity through which a natural person gains an economic benefit on a regular basis counts as professional, expressly including “freelance activity”. The Commission’s FAQ on Article 50 (page as of 24 July 2026) concludes: “In this case, the natural person is considered to be a deployer of that AI system.” At the same time, para. 19 states that use by natural persons acting on behalf of or under the authority of a professional deployer falls within that deployer’s transparency obligations. So if you work to the specifications of a client or an agency, that party bears the obligation; if you decide yourself, on the most natural reading you are the deployer.

The Centre for Protection against Unfair Competition (Wettbewerbszentrale) in Germany also reads the concept of deployer broadly in its guide (version 2.0, as of 29 July 2026): according to the guide, it covers not only large companies but also sole traders, social media professionals and marketing firms that use AI for advertising (own translation). The guide does not address how the role is split between an agency and its client.

When do you become a provider yourself?

Under Article 3(3), a provider is a person or body that develops an AI system or has it developed and places it on the market or puts it into service under its own name or trademark. Anyone who operates ready-made services such as ChatGPT, Adobe Firefly, Midjourney, Canva or Magnific is not. Magnific (Freepik) states this expressly in its terms of use (June 2026 version, section 14.2): “For the purposes of the AI Act, Users of the AI Products act as deployers of AI systems.”

It is different if you adapt a model. Under para. 11 of the guidelines, a company becomes the provider of a new system if it modifies an existing generative system from another provider, for example “with new training data”, and then puts it into service under its own name. If you use such a system yourself for deep fakes, you are both provider and deployer under para. 15: you then owe the machine-readable marking under Article 50(2) and the visible disclosure under Article 50(4). Article 25 of the AI Act, under which distributors, deployers or other third parties can be treated as providers, only concerns high-risk systems (paragraph 1) and plays no role here.

What remains open is from what degree of adaptation para. 11 applies, for example a small amount of additional training on a client’s product photos. Under Article 111(4) (inserted by Article 1 point 39(b) of Regulation (EU) 2026/1744), the transitional period until 2 December 2026 only applies to providers of systems placed on the market before 2 August 2026. If you want to use your own or an adapted model, have the provider question checked separately beforehand.

Which of your images need a label?

Under Article 50(4), only a deep fake needs a label: under Article 3(60), “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 guidelines expressly count consumer goods among objects and realistic AI-generated avatars among persons (para. 113). Whether an image appears authentic is assessed objectively, taking account of context and audience; no intention to deceive is required (para. 114). Colour correction, noise reduction, rescaling, replacing backgrounds for clearly aesthetic purposes and arranging existing products are likely to have only a minor impact in product advertising under para. 116, subject to the individual case.

For typical jobs, this means the following, classified according to the guide Which Images in Your Store Need an AI Label?, which lists all image types with reasons:

The variant follows the question “Was there an original photo?”. According to the Commission’s icon page (page as of 24 September 2026), AI GENERATED is for content fully generated by AI with no human-created elements apart from the prompt, and AI MODIFIED for pre-existing, human-made content that was partially modified with AI, turning it into a deep fake. Official icon files only exist with English wording; the Wettbewerbszentrale recommends checking whether German “KI” is needed instead of “AI” for German-speaking audiences. Details on variants, language and placement: The EU icons for AI images.

Creative in the legal sense is not the same as creative work. Article 50(4) allows a more restrained disclosure for deep fakes that form part of evidently artistic, creative, satirical or fictional works. For product images this usually does not help: works “mainly motivated by functional or technical considerations” are not regarded as creative under para. 120, and content whose nature is exclusively informative or commercial and recognisable as such is excluded under para. 122. Advertising can therefore only count as an evidently creative work in certain specific situations.

What should you hand over to your client?

Whoever the deployer is, the store needs information that only you have. If your client is the deployer, it has to label and is expected to be able to explain how it has done so (para. 148). If you are, your labelling should reach the audience, and that audience sees the images in your client’s store (para. 12). A handover package that covers both cases:

1. List of images with AI involvement: file name, motif and product, tool and model as far as shown, date, source material (own photo, client’s photo, prompt only) and what the AI did (created a person, replaced the background, changed a product detail, retouching only).
2. Your assessment for each image: deep fake yes, no or borderline, with a one-sentence reason, and the proposed variant AI GENERATED, AI MODIFIED or no label. The decision lies with the deployer; your proposal makes it traceable.
3. The original files as they came out of the tool: via the download button in the native format, without screenshots and without conversion, with unaltered metadata. The edited final versions separately.
4. Channel versions, if agreed: separate derivatives with the icon burnt in for ads, marketplaces or newsletters. The original file stays without a burnt-in icon so that suitable versions can be derived for each channel.
5. Rights of persons: in addition to labelling, the rules protecting a person’s own image apply (para. 129). In Germany, portraits may in principle only be disseminated with consent (Section 22 KUG, the German Art Copyright Act); this concerns real models as well as AI persons that recognisably resemble a real person.
6. Faithful product depiction: which features of the goods the image shows and whether the AI changed any of them.

Why the original files matter: a provenance marker (C2PA, IPTC DigitalSourceType) is evidence of AI origin, while a missing marker says nothing. The guidelines encourage everyone who passes content along the chain to preserve existing markings (paras. 16 and 98). Google Merchant Center requires images created with generative AI to contain such metadata and asks that embedded metadata tags such as DigitalSourceType are not removed (AI images in your Google Merchant Center feed). And generators can prohibit removal: the usage policy of Black Forest Labs, provider of the Flux models (last revised 4 August 2026), forbids removing, altering or suppressing any “C2PA Credentials, digital watermarks, or other content provenance signals”.

What belongs in the contract?

The following list is a checklist for the conversation with your client and for review by a lawyer, not a ready-made contract clause and not standard terms. The guidelines themselves name the contract: deployers in production and distribution chains should take proportionate measures so that their labelling reaches the audience, “e.g., via contractual conditions with distributing partners” (para. 12).

How does the label get into the store and the channels?

In the store. Under Article 50(5), the information must be provided “in a clear and distinguishable manner at the latest at the time of the first interaction or exposure”, for each image and to every person who sees it (para. 143), without technical tools and without any action by the viewer (para. 117). Information hidden in menu layers or terms of use is not enough (para. 142). The label therefore belongs on the image wherever it appears in the store: product gallery, collection, home page, banners, blog. The Code of Practice, which only binds signatories, names a user interface overlay that appears on the image as an equivalent alternative to an embedded icon. In the store, an overlay is often the simpler route because the image file stays unchanged; an app such as Easy AI Act Image Labels shows the EU icon this way in the online store and sets no label in feeds, ads or emails.

Outside the store. An overlay does not travel with the image when it goes into a feed, onto a marketplace, into an ad or a newsletter. The Commission’s icon page sums up that the icon must remain visible when content is reshared or downloaded; the Code of Practice only requires best efforts for this. These channels need the platform’s own labelling route or a separate image version with the icon burnt in, where the platform allows it; there, the word variant belongs in the image, not the basic icon alone. What each platform requires or prohibits: AI images on marketplaces and in ads.

If you are the deployer. Then it is your labelling that has to reach the audience. The guidelines expect deployers in production and distribution chains to take “proportionate measures” so that the labelling is displayed clearly to the targeted audience at the point of first exposure, for example through contractual conditions with distributing partners or user experience settings (para. 12). Which audience you have to reckon with depends on the intended distribution channels; you are not expected to consider further dissemination by third parties beyond that (para. 115). Anyone producing for a store knows that the images will appear there and in the agreed channels (our reading). In practice, two routes remain, and they can be combined: you deliver labelled versions, or you agree that your client places the label and give it everything it needs. The guidelines do not say how far “proportionate” goes, for example whether you have to check the implementation in the store.

Which tools write provenance evidence?

Whether an image carries its AI origin with it depends on the tool, often even on the individual model. On the Easy AI Act Image Labels test bench on 1 September 2026, ChatGPT, Google Gemini, Canva and Shopify Tinker wrote a marker into the file; Adobe Firefly only wrote a reference to an online manifest, the Shopify image editor (“Generate background”) wrote no metadata at all, and Midjourney has not been measured. In the Magnific catalogue, 34 of 47 text-to-image models wrote a marker and 13 did not. The table by tool and model is in the guide Which AI image generators write provenance evidence?. For Shopify’s own AI features, see Shopify Magic and the EU AI Act.

Three things follow for your handover:

• The export route matters. The same generation carried a C2PA manifest as a native PNG but no marker as a JPEG from the web interface; a screenshot kept no provenance information (measurement of 1 September 2026). Deliver the file from the download button, not a conversion.
• The store copy usually loses the marker. The original stored in Shopify kept all 34 markers on 1 September 2026; on the copies Shopify delivers to visitors, only 6 of 38 marked images still carried their marker on 30 September 2026. What remains checkable is therefore mainly the original file.
• A marker does not say whether a label is needed. Image editing with AI features can also write a marker: in a public store, Shopify delivered images with a C2PA manifest on 30 September 2026 that had been edited in Adobe Photoshop, with Adobe Firefly named as the AI tool in the file. Whether such an image is a deep fake only your assessment can answer. Conversely, an image without a marker has not necessarily been made without AI.

The free AI image checker shows whether a file carries evidence, without the file leaving your device. A marker never replaces the visible disclosure: under para. 117, deployers cannot rely on the machine-readable marking embedded by the provider under Article 50(2).

When does a label not help?

The AI Act governs the disclosure, not whether the image is permissible. Under Article 50(6), paragraphs 1 to 4 “shall be without prejudice to other transparency obligations laid down in Union or national law for deployers of AI systems”. And under para. 129 of the guidelines, the labelling obligation does not mean that deep fakes that are unlawful under other law may be generated and disseminated; the guidelines name “misleading advertising” as an example. If an image shows the goods differently from how they are delivered, that is a matter of unfair competition law, with or without a label (the following summaries of German-language statutes are our own translations):

• Germany, Section 5 UWG. A commercial practice is misleading if it contains untrue or other deceptive statements about the essential characteristics of the goods, such as their type, design, composition or quality (subsection 2 no. 1). Under subsection 4, statements also include pictorial representations and other arrangements that are intended and suitable to replace such statements.
• Austria, Section 2 UWG. A commercial practice is deemed misleading if it contains incorrect information or is otherwise likely to deceive a market participant about, among other things, the existence or nature of the product or its main characteristics (subsection 1, points 1 and 2).
• Switzerland, Article 3(1)(b) UWG. Anyone who makes incorrect or misleading statements about their goods, or in a corresponding manner favours third parties in competition, acts unfairly. How far the second part covers an agency that creates images for a retailer is not answered by this guide.

Under the guidelines, the deep fake criterion of the AI Act is to be understood independently of the concept of deception in the Unfair Commercial Practices Directive (footnote 32). An image can therefore be labelled correctly and still be misleading. Typical risks in production: the AI smooths fabric that actually creases, makes colours richer than the goods, adds details that do not exist, or shows a fit the product does not have. “Illustrative image” or “image may differ” is no AI disclosure and does not remove the deviation. Such images need correcting, not labelling, ideally after approval by the client, who knows the goods.

Who is liable for what?

This question can only be answered as far as statutes expressly govern it. As of 30 September 2026, there was no publicly documented enforcement case on labelling AI images in online retail, whether an authority proceeding, a court ruling or a published warning letter (tracker for October 2026).

• AI Act. The obligation under Article 50(4) falls on the deployer, and Article 99(4)(g) links the fines framework to “transparency obligations for providers and deployers pursuant to Article 50”. For actors who are neither providers nor deployers, the AI Act neither extends nor affects their responsibility under other frameworks (para. 17). How an agency and its client split the labelling between them is not governed by Article 50; the guidelines refer to contracts for this (para. 12).
• Unfair competition law. It is your client who advertises with the image. In Germany, claims for removal and injunction under Section 8(1) UWG lie against whoever carries out the unlawful commercial practice and, under subsection 2, also against the owner of the business if the infringements are committed by an employee or agent (own translation). In Austria, under Section 18 UWG the owner of a business can also be sued for an injunction if the act was committed in the operation of the business by another person (own translation). Whether an external agency falls under these rules in a given case, and when the agency itself can also be held liable, is answered neither by the wording of the statutes nor by this guide.
• Between you and your client. What you owe is determined by the contract under the applicable contract law. The terms of the tools also set limits for you as a user: Magnific expressly places the disclosure of deep fakes on its users (“users must clearly and visibly disclose”, section 14.2), and Black Forest Labs prohibits content that falsely represents events, things or people in a manner likely to deceive a reasonable viewer “without appropriate or lawful disclosures”.

More on supervision, the fines framework and the risk of warning letters is in the guide Who enforces the EU AI image labelling rule in your store?.

How do you document your AI images?

Deployers that have not signed the Code of Practice “are expected to demonstrate how they have complied with their obligations under Article 50(2), (4) and (5) AI Act through other adequate means” (para. 148). The authorities will likely need more detailed information from them; deployers, too, may receive requests, expressly “with regard to their labelling practices under Article 50(4) AI Act”. As an example, the guidelines name a comparison of one’s own measures with those of a code of practice assessed as adequate. If you are the deployer, this applies to you directly; if your client is, it needs your records. One line per image is enough, with:

• job, client, file and place of use (store, channel),
• tool, model, date and source material,
• type of AI use,
• assessment deep fake yes or no with a short reason, and the chosen variant,
• who decided on the use of AI, supported by the briefing, approval or selection,
• original file with marker yes or no, with the date checked,
• handover: date, recipient, versions delivered.

The line on the decision is the most important. Under para. 14, the role depends on who decided whether and how, and later only briefings, approvals and selection records will show that. What a retailer should record itself is in the enforcement guide linked above.

Does this also apply to agencies in Switzerland?

The AI Act applies to deployers established in the Union (Article 2(1)(b)) and to deployers in third countries such as Switzerland “where the output produced by the AI system is used in the Union” (point (c)). Under para. 13 of the guidelines, this is the case where the deployer itself foresees dissemination in the Union, for example by directing or authorising it. Para. 14 gives the example of a company in a third country that uses AI to generate a deep fake of a celebrity for an advertisement displayed in the Union. On that basis, a Swiss agency that produces images as the deployer for stores with customers in the EU is likely to be covered (our reading). What applies under Swiss law itself: Labelling AI images as a Swiss shop.

What is still open?

• The line under para. 14: the guidelines do not say from when briefings, reference images or choosing between AI variants amount to deciding on the “how”.
• Two deployers: the guidelines leave open whether an agency working on behalf of and under the control of its client is also a deployer itself.
• Reach of para. 12: how far the “proportionate measures” go when an agency is the deployer, for example whether a contract clause is enough or the agency has to check the implementation in the store, is open.
• Provider threshold: from what degree of modification of a model an agency becomes a provider under para. 11 is open.
• Binding force: under para. 5, the guidelines are non-binding; only the Court of Justice of the EU can interpret the AI Act authoritatively. No court ruling on labelling AI images in online retail was known as of 30 September 2026.

In short

If you create images with AI for retailers, first clarify the roles: if the client decides whether and how AI is used, the client remains the deployer; if it only orders the result, on the most natural reading the role lies with you. You only become a provider with a modified model under your own name. Either way, the store needs a list of the AI images from you with the tool, your assessment and the label variant, plus the original files with unaltered metadata; roles, labelling, deliverables and faithful product depiction belong in the contract. A label does not make a misleading image permissible, and briefings and approvals will later show who decided.