EU AI Act and AI product images: provider and seller duties (2026)

The EU AI Act separates a tool provider’s machine-readable marking duties from a business user’s disclosure duties for deepfakes. Article 50 generally applies from August 2, 2026; the December 2, 2026 transition concerns only Article 50(2) provider duties for qualifying systems placed on the market before August 2.

This practical reading was checked on October 21, 2026 against the Commission’s published guidance and legal text. It is general information, not legal advice or a certification of any tool’s compliance. Have a qualified adviser assess the facts of your business when the classification or obligations are uncertain.

Which part of Article 50 matters for product images?

For product-photo workflows, two provisions deserve separate treatment. Article 50 addresses both the origin markings a system provider must enable and the disclosure a business user must make for content that constitutes a deepfake.

This split changes the questions you ask a supplier. “Do your downloads contain AI metadata?” is useful. It is different from “What must my customer see when I publish this particular image?”

Who is the provider, and who is the deployer?

The Act’s Article 3 definitions distinguish the entity developing, or having developed, an AI system and placing it on the market or putting it into service under its name from the entity using a system under its authority. The deployer definition excludes personal, non-professional use.

A clothing shop using a third-party image generator for its business should therefore assess the deployer role. It should not assume that purchasing a subscription transfers every publishing obligation to the software company. A business offering its own branded AI service needs a separate assessment of whether it is a provider.

Record the real arrangement rather than choosing a label that sounds convenient. Who offers the system? Whose name appears on it? Who decides what the output depicts and where it is published? Who approves the finished image?

The Commission FAQ also discusses professional use and contractors working under a business’s authority. Outsourcing a campaign is not, by itself, a complete answer to responsibility. Put the image-review and publishing tasks into the brief you give the agency.

What exactly changes on December 2, 2026?

The general Article 50 date is August 2, 2026. The narrower transition concerns providers of qualifying generative systems placed on the market before that date and their Article 50(2) marking and detection obligations.

The operative condition appears in Article 111(4), added by Regulation (EU) 2026/1744. It is a system-and-role condition, not a general holiday for retailers.

That means a seller should not substitute any of the following dates for the legal test:

• The date the shop bought its subscription.
• The date a contractor began the campaign.
• The date the seller opened a marketplace account.
• The date someone downloaded an image.

Those dates can be useful evidence about a project, but none independently establishes when the relevant AI system was placed on the market.

Ask the provider to identify the system, explain whether it relies on the transition, and state its implementation timetable. Keep the answer with your production records. Assess your own publishing duties separately instead of treating the supplier’s timetable as the deadline for every image.

Does an AI product photo automatically count as a deepfake?

No universal answer follows from the fact that AI was used. Article 3(60) considers the resemblance to existing persons, objects, places, entities or events and the false appearance of authenticity or truthfulness. Applying that definition to a particular product image is a contextual assessment.

Compare three possible briefs. In the first, an editor cleans dust from an unchanged photograph. In the second, a generator places a recognisable real item in an invented location. In the third, a generated person appears to wear a recognisable garment in what looks like a real shoot.

Those examples raise different questions. What has been changed? What does the image communicate? What would the intended audience understand it to show? Would a product scene imply a real event, endorsement or property of the item that the seller cannot substantiate?

These are prompts for review, not official rulings on those examples. Do not label all product photos deepfakes, and do not decide that ordinary ecommerce imagery is always exempt.

For an uncertain case, give an adviser the original, final image, edit history, proposed caption and actual placement. A detached thumbnail rarely captures everything needed for a useful assessment.

What does the standard-editing exception actually cover?

Article 50(2) includes an exception concerning assistive standard editing and changes that do not substantially alter the input or its semantics. That wording should not be converted into “every background edit is exempt.”

The operation’s name is not enough. A button called “clean up” might remove a speck, reconstruct a missing product edge or invent a large part of a garment. A button called “background” might retain a plain backdrop or create a convincing scene suggesting circumstances that never occurred.

Keep a before-and-after pair and describe the actual change in plain language. “Removed a small sensor spot” is more informative than “used enhancement.” “Reconstructed the left sleeve and replaced the setting” gives a reviewer a different factual starting point.

The exception is written into the provider-marking provision. It should not be casually expanded into a blanket exemption from marketplace rules, consumer-facing accuracy, or any applicable disclosure duty. You still need to assess the final content and its use.

Is IPTC metadata enough to satisfy the law?

Do not treat one tag as a complete compliance certificate. Article 50(2) calls for technical qualities including effectiveness, interoperability, robustness and reliability, with feasibility and other specified considerations. The legal assessment is broader than whether a keyword can be found in an export.

Useful supplier questions include:

• Which outputs receive markings, and where is the mechanism documented?
• What happens when a file is resized, converted or downloaded in another format?
• How can a recipient detect or inspect the marking?
• What limitations are known, and how are they communicated?
• Which obligation is the supplier addressing, and which decisions remain with the publisher?

CatalogCut Stage downloads include IPTC DigitalSourceType metadata and, for on-model images, Amazon’s contains-synthetic-performer keyword. The Stage workflow describes its generation and product-checking process. These are product facts; they do not establish a legal conclusion about every system, export chain or seller use.

If you optimize an image, inspect the output after that final step. The image optimizer’s metadata guidance explains which disclosure fields it preserves. A description of intended preservation is still different from checking the exact file your campaign will publish.

When does a seller need a visible disclosure?

Where the deployer deepfake-disclosure duty applies, the Commission explains that people must be able to perceive the disclosure at first exposure. A hidden field that requires a metadata reader is insufficient on its own.

Plan the disclosure alongside the placement. Ask how a person encounters the image in a listing gallery, a social ad or a cropped preview. A note visible only on a separate landing page should not simply be assumed to solve the audience’s first-exposure question.

This can create a practical conflict with a marketplace’s image rules. The solution is not to hide a legally required disclosure to preserve a clean main image. Consider a different image, a different placement or specific advice about a compatible presentation.

The marketplace image and AI rules reference distinguishes platform requirements. For example, a marketplace metadata instruction and an EU disclosure assessment answer different questions. Satisfying the first does not establish the second.

What should you change in your production workflow?

Create a small record for each approved final image. This is a practical review aid, not a claim that the Act mandates this exact record format.

Include the source photograph, a useful description of the generation or edits, the provider and system identifier, the creation date, and the intended channel. Record who checked product fidelity and who decided the disclosure treatment.

Keep machine-readable marking checks and audience-facing disclosure decisions as separate entries. Otherwise, “tagged” can become an ambiguous status that hides unfinished work.

For example, a campaign manager might approve the product details but leave an unresolved question about a realistic generated person. Mark that question as unresolved rather than turning the image into a scheduled post. A technical reviewer might confirm a metadata field while lacking the context to assess the advertisement’s message. Record that limit too.

Add this process to your product-photo style guide so a freelancer or colleague can follow it without reconstructing the decision from chat messages. Preserve the approved final file alongside the working original; later export changes should trigger another file and placement check.

What to do next

Choose one live image workflow and identify the provider, final publisher, export chain and destination. Resolve role, date and disclosure questions with the right adviser, then use the AI image metadata tagger to inspect the Amazon and Google fields as one specific part of the file review.