EU Deepfake Rules Hit Retail: Eurocommerce Demands Exemption
A sofa staged in a sunlit living room, rendered entirely by Stable Diffusion. Is that a deepfake? On August 2, the EU AI Act's transparency rules for synthetic content start applying, and Eurocommerce — the trade body…
EU Deepfake Rules Hit Retail: Eurocommerce Demands Exemption
A sofa staged in a sunlit living room, rendered entirely by Stable Diffusion. Is that a deepfake? On August 2, the EU AI Act’s transparency rules for synthetic content start applying, and Eurocommerce — the trade body covering Amazon, H&M, Inditex, and Ikea — wants a written answer before its members start labelling every catalogue render as potentially deceptive media.
The dispute exposes a drafting weakness compliance teams have flagged since the trilogue. Article 50(4) of the AI Act defines a deepfake as AI-generated image, audio or video content “resembling existing persons, objects, places, entities or events” that would falsely appear authentic. By the literal text, a synthetic product shot of an IKEA bookcase qualifies. It resembles a real object, and consumers would treat it as a photograph.
The retail problem is a definitions problem
Eurocommerce’s position to the Commission argues commercial product imagery has nothing in common with the fraud and non-consensual material the deepfake clause was written to address. Their members already operate under the Unfair Commercial Practices Directive 2005/29/EC, which prohibits misleading advertising regardless of how the image was made.
The Commission’s own AI Office has not clarified the boundary. ENISA’s 2024 threat landscape lists deepfakes under disinformation and fraud — not retail catalogues. Yet the statute makes no carve-out.
ai_act_disclosure_check:
content_type: product_render
resembles_real_object: true
context: advertising
art_50_4_applies: true # literal reading
ucpd_safe_harbour: unclear
default_action: label_as_ai_generated
cost_per_sku_eur: 0.04For a retailer with 400,000 SKUs refreshed quarterly, that default action is not trivial. The labelling, the audit trail, the dispute handling all need a controller. This is where compliance teams have been quietly building tooling that maps each asset class to its specific Article 50 obligation — the same architecture pattern we covered in our risk classification breakdown.
How the classification decision actually runs
The Commission has signalled some flexibility. Recent AI Act implementation notes acknowledge gaps; our tracker on the broader AI Act timeline follows the moving deadlines. None of it resolves the retail question.
Where CAI Technology sits on this
The Eurocommerce request is reasonable; the legislative fix is not. Carving out “advertising” creates a worse loophole than the one it closes — political ads, financial promotions and pharmaceutical marketing would all queue up for the same exemption. The cleaner path is what the AI Office can already do under Recital 134: issue codes of practice that distinguish deceptive resemblance (a fake person, a fabricated event) from productive synthesis (a styled render of a real product). Retailers get certainty without rewriting the statute.
We’re watching how this lands for retail clients running image pipelines through Lexnomia’s compliance layer. If you want a walk-through of the disclosure matrix for your SKU catalogue, our note on the fine-tuning and content obligations retailers miss is the place to start.