Reading time: 8 minutes
Since August 2, 2026, the transparency obligations under Article 50 of the EU AI Act have been in effect. Shortly before this, the EU Commission published its final guidelines and voluntary labeling symbols.
Crucial for Shopify merchants: Not every image created or edited with AI requires a visible notice. The decisive factors are content, realistic effect, and context of use.
Quick & Compact
✓ There is no general obligation for a visible label on every AI image.
✓ Visible notices are primarily required for AI-generated or manipulated deepfakes.
✓ In addition to people, deepfakes can also depict products, places, objects, and events.
✓ Technical watermarking is primarily the responsibility of the provider of the AI system.
✓ Product images and advertising motifs must therefore be evaluated based on their potential to deceive.
What has been in effect since August 2, 2026
Article 50 distinguishes between providers and deployers of AI systems. The provider develops a system or places it on the market under their own name; the deployer uses it professionally under their own responsibility.
A Shopify business that creates product motifs or ads using an image generator is generally considered a deployer. This remains the case even if employees, agencies, or freelancers operate the system on behalf and under the control of the company.
Technical marking is not the same as a label
Providers of generative AI systems must fundamentally mark synthetic image, audio, video, and text outputs in a machine-readable format and ensure their detection. This technical layer can, for example, be based on metadata or other provenance signals.
The obligation is initially directed at the system provider, not across the board to every merchant who uses its output. Exceptions may include, among other things, assistive functions for standardized editing that do not substantially change the input data or its meaning.
Important
A machine-readable marking does not replace a visible or perceptible notice for customers in the case of a deepfake that requires labeling. Both layers serve different purposes.
When an AI image must be visibly labeled
For merchants, Article 50 Paragraph 4 is particularly relevant. According to this, it must be disclosed if an AI system generates or manipulates image, audio, or video content that constitutes a deepfake.
The term is broader than is often assumed. It includes content that resembles existing or plausibly existing persons, objects, places, entities, or events, and appears falsely authentic or true to the viewer.
The three crucial questions
Does the motif highly resemble an existing or plausibly existing person, product, place, company, or event?
Could the depicted object actually exist or have plausibly existed?
Could the audience mistakenly understand the image in this specific context as an authentic photograph or a truthful representation?
According to the Commission's guidelines, a deepfake is only likely when these criteria come together. The mere involvement of an AI tool is not sufficient.
Product images, campaigns, and ads in a practical check
Use case | Classification | Visible notice |
|---|---|---|
Minor retouching, background removal, or color correction without changing the product message | Can be considered standardized editing; generally not a deepfake | Usually no |
AI-generated image of a product physically for sale, which looks like a real photo | Can resemble an existing object and appear falsely authentic | Individual assessment; in case of realistic apparent authenticity, likely yes |
Product is enhanced with AI with features, dimensions, or components | High relevance for deception; deepfake classification may be likely | Plan with a notice as a rule |
Recognizably surreal or fantastical campaign motif | Audience mostly does not expect an authentic representation | Often no, as long as no real situation is simulated |
A real influencer, employee, or celebrity is synthetically placed into a scene | Typical deepfake case if the depiction looks real | Yes |
An authentic room photo is virtually furnished with products | Can be considered a partially AI-manipulated deepfake | As a precaution in case of a realistic effect, yes |
Social ad shows a fictional but realistic-looking customer event | Can depict a seemingly true event | Deepfake assessment required |
Product images in particular are a gray area. Customers normally expect to see the product actually offered, its color, proportions, and essential properties on a product detail page.
The more a synthetic image exploits this documentary expectation, the more likely it is to appear falsely authentic. A notice of AI use also does not release you from other requirements for accurate product representation or advertising.
Practical Check
Do not just ask: Was AI used? Ask: What facts does an average customer believe about the product, person, or usage situation based on this motif?
What applies to campaign motifs and social ads
The format used does not change the assessment. A deepfake fundamentally requires labeling, whether it is played out in the Shopify store, in a newsletter, on a landing page, or as a social ad.
The disclosure must be clear, distinguishable, and accessible at the latest upon first contact. A notice visible only after a click or exclusively machine-readable metadata would likely not be sufficient for a deepfake.
The EU provides optional symbols for fully AI-generated and partially AI-modified content. Their use is voluntary and does not on its own prove compliance; a comprehensible text like AI-generated or partially modified with AI may additionally make sense depending on the medium.
Deepfakes in creative works
For obviously artistic, creative, satirical, or fictional works, transparency is not entirely omitted. However, it may be implemented in such a way that it does not impair the presentation and enjoyment of the work.
This limitation may be relevant for highly staged campaigns. However, whether classic product advertising already qualifies as an obviously creative work cannot be answered generally; in the case of realistic-looking claims about a product, restraint is advised.
Correctly distinguishing editorial content
The deepfake assessment still applies to editorial images. A human final review does not automatically exempt a realistic AI image from the obligation to disclose.
The specific exception for human review and editorial responsibility applies to AI-generated or manipulated texts published to inform the public on matters of public interest. This can include, for example, consumer safety, health, the environment, politics, or economic developments.
A mere spelling or grammar check is not enough. What is required is a substantial review by expert individuals or real editorial control with the authority to change or reject content and check sources; in addition, an individual or organization must bear legal responsibility for the publication.
Legal Uncertainty
Whether a specific product motif appears as a deepfake depends heavily on the context, target audience, and their expectations. This practical guide does not replace an individual assessment and does not constitute legal advice.
What store operators should check now
✓ Record which product images, campaign motifs, and ads were created or substantially modified using generative AI.
✓ Document the original, the system used, the editing, and the responsible approver for each asset.
✓ Evaluate realistic motifs based on the three deepfake criteria instead of using a blanket AI label.
✓ Define clear notices for fully generated and partially manipulated content.
✓ Place required notices at the point of first visual contact, and consider downloads and further distribution.
✓ Check whether exports, agency processes, and technical image processing remove existing origin signals.
✓ Establish a substantial human review for texts on topics of public interest.
A streamlined approval process
Record origin: Human-created, assistively edited, partially AI-modified, or fully generated.
Check reality effect: Could the motif be understood as a real photo of a product, person, place, or event?
Assess deepfake risk: Document resemblance, existence, and apparent authenticity.
Determine notice: No notice, voluntary transparency, or mandatory disclosure.
Control delivery: Review shop, ads, newsletters, downloads, and agency variants together.
The Commission and the European AI Board have evaluated the voluntary Code of Practice as an appropriate tool to support compliance. However, signing it or using the EU symbols is neither mandatory nor a final proof of compliance.
Thinkideas View
Shopify merchants should not slap a blanket badge on every AI motif. A documented approval process makes more sense, consistently labeling realistic product depictions and synthetic testimonials, while not unnecessarily burdening harmless standard edits.
Conclusion
Since August 2, 2026, the EU AI Act does not require visible labeling of every AI image. What matters most is whether a motif appears falsely authentic or true as a deepfake.
Start with an inventory of your active assets and prioritize realistic product images, virtual scenes, motifs depicting people, and social ads. You should document borderline cases and have them legally reviewed.
