General information, not legal advice.
In September we flagged the intersection nobody was covering: fashion writes its product copy with generation templates, and two bodies of European law now read that copy closely. This is the promised deep dive into the AI Act half of the problem: when does AI-written text legally require its own disclosure, why most fashion copy escapes it, and which three kinds of claims pull it back into scope.
The text rule, plainly
The AI Act's disclosure duty for text is deliberately narrow. It covers AI-generated or AI-manipulated text published with the purpose of informing the public on matters of public interest. Not all AI text; not marketing in general; text that does the job of journalism or public information. And even inside that scope there is a built-in exception: the duty falls away where the AI-generated text has undergone human review or editorial control, and a person or organisation holds editorial responsibility for its publication.
Read together: the rule targets machine-written public information published on autopilot, with no human answerable for it. That is the mischief the law is aimed at, and it explains both why your product pages are normally safe and exactly how they stop being safe.
Why ordinary product copy is outside
"Relaxed-fit organic cotton shirt, garment-dyed, mother-of-pearl buttons" is commercial speech describing a product. It informs a buyer, not the public, and fit, fabric and care instructions are not matters of public interest in the legal sense. Generate a million of these with templates and the AI Act's text-disclosure duty stays out of the room. Whatever else those million descriptions must respect, and the green-claims rules are the whatever else, they do not need an "AI-written" disclosure.
The three claim types that pull copy toward scope
The boundary moves when the copy stops describing the garment and starts informing the public about the world. The Guidelines treat sustainability, health and similar topics as public-interest matters, and where content mixes the commercial and the informative, the informative character prevails, the same logic that decided the deepfake question for images. Three claim families do the pulling:
- Environmental claims. "This fabric saves 3,000 litres of water per garment." "Our supply chain is carbon neutral." The sentence is no longer about the shirt; it is a factual claim about environmental impact, exactly the territory the Guidelines read as public interest. AI-generated sustainability pages, impact reports and "our planet" sections sit furthest into scope.
- Health and wellbeing claims. "Antibacterial fibres." "UV-protective." "Hypoallergenic." Health information is the other named public-interest category, and performance-of-the-body claims drift the same way environmental ones do.
- Supply-chain and ethics claims. "Made in certified fair-wage facilities." Claims about labour conditions and sourcing inform the public about conduct in the world, not about the drape of the fabric.
Notice the pattern: the more a sentence would fit in a press release or a news article, the further it sits from "product copy" and the closer to the disclosure duty.
The exception that saves you, if you build it
Here is the practical centre of this post. The human-review exception means the disclosure duty does not attach where the text passed human editorial review and someone holds named editorial responsibility. For a fashion brand, that converts a legal question into a workflow question: does claim-bearing copy ship straight from the template, or does it pass a person whose name is on it?
To be usable, the exception has to be real and demonstrable, not decorative. That means a defined review step for the claim-bearing zones below, a named editorial owner (a role that pairs naturally with the compliance owner your readiness check asked about), and a record that the review happened. A rubber stamp that approves a thousand generated pages per second is the kind of "review" an authority is entitled to look through.
The three zones, one table
| Zone | Example copy | What applies |
|---|---|---|
| Zone 1 · Pure product description | Fit, fabric composition, care, styling suggestions | No AI text disclosure duty. Green-claims rules only if environmental vocabulary sneaks in, which is a prompt-guardrail problem. |
| Zone 2 · Claim-bearing product copy | Water savings, carbon claims, antibacterial or UV performance, fair-wage sourcing, on a product page | Substantiation duty for the claim itself (consumer law), and drift toward the AI text rule: route through human editorial review with named responsibility. Two laws, one sentence. |
| Zone 3 · Informational content | Sustainability pages, impact reports, guides and articles generated with AI | Closest to squarely in scope. Human editorial review and named responsibility are the working control; without them, AI disclosure. Treat autopilot publication as not an option. |
The workflow, in four moves
- 1. Sort your copy surfaces into the three zones. Most catalogues are 95 percent Zone 1, and knowing that is already relief.
- 2. Guard the prompts: the restricted green vocabulary stays out of Zone 1 templates entirely, the one-paragraph fix from September.
- 3. Gate Zones 2 and 3 behind human editorial review with a named owner and a record. This one step works double: it grounds the AI Act exception and it is where claim substantiation gets checked.
- 4. Keep the evidence file per claim: the substantiation behind every environmental, health or sourcing statement that survives review. No file, no claim.
Text was always the quieter half of the transparency rules, which is exactly why it will produce the more surprising cases. The image half of the framework, and everything else this series has covered, lives in the Article 50 guide and on the compliance page.
Every image exported from Modelia carries signed provenance and an invisible watermark by default, and anyone can check any image, free, no account, in our public C2PA checker. The full picture, rules and industry data, is in The State of AI-Generated Content in Fashion 2026.
