EU AI Act Article 50 for ecommerce creative: a practical 2 August 2026 disclosure checklist for AI-generated product images, video ads, virtual try-on and campaign content
A practical Article 50 checklist for EU-facing AI product images, video ads, virtual try-on, campaign content, and customer-facing AI.
Lamina Team
Product Team @ Lamina
What will EU AI Act Article 50 require of ecommerce creative from 2 August 2026?
| Metric | Value | Source |
|---|---|---|
| Transparency obligations apply from | 2 August 2026 | European Commission |
| Code of Practice signatories | ~190 | European Commission |
| Maximum company fine | €15 million | European Commission |
| Maximum turnover-based fine | 3% | European Commission |
| Maximum EU institution fine | €750,000 | European Commission |
“We are taking an important step towards AI that people and businesses can understand and trust, and whose benefits are widely shared across our society,”
Article 50 of the EU AI Act does not require a visible label on every ecommerce asset that AI touched. It draws a line between duties for an AI-system provider and duties for the retailer, brand, agency, or marketplace publishing the creative. The European Commission states that Article 50 applies from 2 August 2026.
For creative teams, start with one question: is the published asset a deepfake? That means an AI-generated or AI-manipulated image, audio, or video resembling an existing person, object, place, entity, or event, which falsely looks authentic or truthful in context. A generated packshot does not automatically qualify. A realistic product demonstration implying that a feature, setting, or event was real may qualify.
Make this part of asset review, not a design tax. Keep producing product scenes, virtual models, try-on outputs, and short-form ads, then make the disclosure call at approval, when the team can judge the actual viewer impression and placement.
| Metric | Value | Source |
|---|---|---|
| Article 50 application date | 2 August 2026 | digital-strategy.ec.europa.euas of 2026-07-29 |
| Provider requirement for synthetic image, video, audio, and text outputs | Machine-readable marking and detectability | digital-strategy.ec.europa.euas of 2026-07-25 |
| Deployer disclosure trigger | AI-generated or manipulated media that falsely appears authentic or truthful | digital-strategy.ec.europa.euas of 2026-07-20 |
| Standard-editing scope limit | Assistive editing that does not substantially alter input data or its semantics is excluded from provider marking | digital-strategy.ec.europa.euas of 2026-07-20 |
| Legacy-content treatment | Content generated before 2 August 2026 does not need retroactive labelling | digital-strategy.ec.europa.euas of 2026-07-25 |
Do online retailers have to label AI-generated product photos?
An online retailer needs a human-facing disclosure for an AI-generated or manipulated product image only if it meets the Article 50 deepfake test in the context where it is published. The Commission’s definition rests on realistic resemblance and a false appearance of authenticity or truthfulness, not AI use by itself.
That line matters on the PDP. AI can create a new background, model, or visual concept while keeping the item truthfully represented; still, review whether the finished image makes a factual-looking claim a shopper could wrongly take as real. Put the call in the asset record. Do not paste one label across the whole catalog.
Keep the vendor’s output marker separate from the retailer’s disclosure. The Commission gives providers of generative systems a machine-readable marking duty; the deployer’s deepfake duty is about informing the person seeing the content.
2 August 2026 Article 50 disclosure checklist for ecommerce creative
Create one asset register for EU-facing work
Include product-detail images, lifestyle scenes, on-model assets, virtual try-on output, video ads, social posts, marketplace media, campaign pages, avatars, and chat experiences. Log the tool, source inputs, prompt or brief, human edits, owner, publication date, and EU countries or audiences reached. For each workflow, identify whether the business is a provider, deployer, or both.
Screen every published image, video, and audio asset for deepfakes
Ask three things: did AI generate or manipulate it; does it resemble an existing person, object, place, entity, or event; and would it falsely look authentic or truthful to the intended viewer in that placement? If the answer is yes to all three, plan the disclosure before publication. Judge the finished asset, not the software category.
Put disclosure where viewers first see it
If the deepfake screen calls for disclosure, use a clear, distinguishable, understandable human-facing notice no later than first exposure. On a reel, that may mean an opening on-screen label; for a display asset, an on-asset or adjacent notice is easier to defend than an explanation buried on a detail page. Machine-readable marks alone do not meet the deployer-facing disclosure described in advertising-law analysis.
Collect provider-marking evidence on its own track
Ask every generative-tool vendor how outputs receive machine-readable marking and how an output can be detected as AI-generated or manipulated. Keep export records and vendor documentation. This sits apart from the retailer’s public-disclosure call, and gives procurement a concrete test when it reviews new tools.
Classify routine editing before calling it synthetic media
Record whether the tool carried out assistive standard editing or materially changed what the shopper sees. The Commission says provider marking does not apply where AI performs standard editing or does not substantially alter the input or its semantics. Do not log colour correction, cropping, denoising, and resizing as generative replacement. A changed product feature, fabricated scene, or realistic event needs tighter review.
Review virtual try-on when it is reused publicly
Assess private shopper-facing try-on output separately from a public campaign asset. A closed interaction using a customer’s own image has a different exposure context; reuse in a paid ad, social post, or marketplace listing needs a fresh deepfake assessment. Check product accuracy, data protection, intellectual-property rights, and personality rights too. An Article 50 label does not resolve those issues.
Keep human editorial control over public-interest campaign text
Article 50 is not a broad rule requiring an AI-copy label on every product description. Its text obligation is narrower: AI-generated or manipulated text published to inform the public on matters of public interest, where it has not received human review or editorial control and has no responsible editor. Assign named review to health, safety, sustainability, and similar claims.
Disclose AI interaction in customer-facing agents when required
For a chatbot, AI sales agent, or avatar built to interact directly with people, assess whether users are told they are interacting with AI, unless that is already obvious from the circumstances. Check the opening message. Check mobile too, not just the desktop design.
Keep the proof, then re-review reused assets
Save the asset decision, approval, first-exposure screenshot or video capture, provider-marking evidence, editorial-review record, and publication date. Review again when an asset shifts from private try-on to public advertising, organic post to paid media, or one country audience to another. The Commission calls its Transparency Code of Practice voluntary. Article 50 requirements are legal obligations.
Separate archived material from legacy content used again
Do not automatically relabel every historic file. The Commission says content generated before 2 August 2026 does not need retroactive labelling. Record whether legacy material is simply held in an archive or materially republished or repurposed after the application date, then document the call.
How should you disclose an AI deepfake in a video ad or social campaign?
Disclose a qualifying AI deepfake in a video ad or social campaign with a prominent, understandable human-facing notice at or before the first view. A hidden machine-readable marker helps provider compliance and traceability, yet it does not by itself meet the deployer disclosure described in the advertising-law analysis.
Fit the format to the medium. A short video needs a legible on-screen treatment a scrolling viewer can actually encounter; a static placement needs a notice tied to the creative, rather than a far-off policy page. Review labels asset by asset. Duration, crop, placement, and the claim implied by the scene can all change what viewers understand.
The European Commission’s voluntary icons may help you set a consistent internal pattern, though they are not mandatory. Consistency speeds approvals. It does not replace the underlying assessment.
Are AI retouching and standard product-image edits covered by Article 50?
Routine AI-assisted edits do not automatically fall under Article 50’s provider marking rule if they are standard editing or do not substantially alter the deployer’s input data or its semantics. The Commission states that boundary explicitly. Creative operations can use it to separate ordinary correction from material synthetic change.
Use a plain internal rule: log what the tool changed and whether it changes a shopper’s understanding of the product or scene. Denoising an image differs from generating a nonexistent product detail, replacing a product demonstration’s setting, or making a fabricated event look documented.
The standard-editing exception does not remove the deployer’s contextual deepfake review. The team still has to inspect the published result and decide whether it falsely appears authentic or truthful.
What evidence should an ecommerce team keep for Article 50 decisions?
Keep a dated asset register, the deepfake decision, proof of first-exposure disclosure when required, vendor-marking evidence, editorial approvals, and reuse history. That makes a fast creative workflow auditable without forcing identical treatment on every generated image.
The evidence file should show who made the call, which source material was used, what AI changed, where the creative ran, and whether anyone later adapted it. Article 50 analysis turns on how the asset is presented, not merely its filename or generation tool.
Set a re-review trigger in the DAM or campaign-approval system. An image made for a private virtual try-on session can carry a different compliance profile once it becomes a public paid ad.
What is the practical Article 50 call for ecommerce creative teams?
Keep AI creative production moving, while requiring a documented deepfake review for every EU-facing published asset that could falsely appear authentic or truthful. Do not add a visible label simply because a generative tool appeared somewhere in the workflow.
Build review into the handoff from generation to publishing. A human art director or approver should check product truthfulness, realistic depictions, placement, and first-view disclosure when the deepfake threshold is met. A weak brief or unsupported product claim still needs fixing before release.
This checklist covers Article 50 operationally and is not legal advice. For high-stakes campaigns involving realistic people, regulated product claims, public-interest messaging, or cross-border distribution, have qualified counsel review the final creative and disclosure treatment.
