EU AI Act Article 50 labelling: 11 AI images labelled with the EU icon and zero gaps in the compliance audit for a Polish FMCG brand

Overview

The ClientThe client is a Polish FMCG brand selling nationwide, with a WordPress product website in Polish and English (22 pages, 149 images in the media library).

The brand used AI-generated images in the most visible parts of its website: main banners, product sections and social sharing cards. When Article 50 of the EU AI Act came into application on 2 August 2026, the board decided to label every AI image, including those the regulation did not require to be labelled retroactively. ROI & Shine designed and implemented labelling in two layers: one visible to consumers (the official European Union icon with a caption) and one readable by machines (an AI declaration written into the image file). The audit of 31 August 2026 confirmed complete labelling across all pages, with no loss to the brand's visual quality.

The Challenge

In FMCG, a product image is a promise made to the consumer. An AI image that shows the packaging differently from what sits on the shelf is a risk of misleading the customer, not just a brand issue. The client faced four obstacles:

• One file, many places of use.The same image appeared at once as an image tag, in a set of resolutions for different screens, as a section background, in link previews and in page scripts. A manual file list would miss some of them, and every unlabelled use is a compliance gap.
• No certainty about what was made with AI.Supplier declarations did not always match the metadata in the files. Without evidence, there is no reliable way to tell a generated image from a modified one or from a regular photograph.
• The crop changes with the screen.Banners are automatically cropped to fit the screen width. The sports section banner lost about 32% of the image at the bottom, and the product pairing section lost 42% at 1440 px. A label placed in the corner of the file simply disappeared.
• The regulation sets no dimensions.The EU requires a "clearly visible" label but defines no minimum size. A label that is too small defeats the purpose of the rule; one that is too large draws attention away from the product.

The Strategy

We treated labelling not as stamping an icon onto images, but as a compliance process that has to survive every future change to the website.

Our strategy rested on three pillars:

• Evidence first, decision second.We established the origin of every image from the metadata stored in the file, not from the team's memory. Before any code was written, the client received a decision document with options and costs, and confirmed the choice in writing.
• Product accuracy before labelling.An icon cannot fix an image that misrepresents the product. We compared every image with the real packaging: label, colour, volume and variant description.
• Numbers from measurement, not intuition.We set the label's size, contrast and position through legibility tests at three screen widths. Every change to the media is checked by an audit, so the label does not vanish with the next photo swap.

The Solution

ROI & Shine implemented two-layer AI content labelling that runs directly on the client's website and in the process of creating new images:

1. 1

   Inventories every use

   A scanner went through the website code and recorded 419 entries in a register: images, resolution sets, style backgrounds, link previews and URLs in scripts. This let us find every place where an AI image was visible, including those that never show up in the media library.

2. 2

   Verifies provenance

   Our own C2PA metadata reader, which needs no external tools, detected 31 files carrying a signed AI declaration. Each image was classified as AI-generated, AI-modified or photograph. Along the way we caught an image showing the wrong product variant description on the label. A product range image with blurred label text was locked against zooming, and its file size dropped from 3,012 kB to 199 kB.

3. 3

   Labels on screen

   A WordPress plugin overlays the official EU "AI GENERATED" icon with the caption "Material generated or significantly modified using AI" (in Polish on the Polish version). The label is anchored to the visible crop rather than to the file, so it stays in place on every screen and in print. On social sharing cards the label is embedded in the image, because no script runs there.

4. 4

   Marks the file and keeps it compliant

   In our content factory, every generated image receives the IPTC entry DigitalSourceType = trainedAlgorithmicMedia in the file: once right after generation and again just before publishing, without re-encoding the pixels. Google Merchant Center has required this same entry for AI-generated product images since February 2024. An audit at 390, 768 and 1440 px and 12 automated tests catch any shifted or missing label.

The Impact

The brand entered the new regulatory regime within the first month of its application, with complete documentation it can present to a supervisory authority, and with no compromise on the look of its website.

How to implement Article 50 labelling: lessons from this project

file,type_of_use,page,origin,evidence,label
banner-sport.webp,css-background,/sport,AI GENERATED,C2PA,yes
banner-sport.webp,link-preview,/sport,AI GENERATED,C2PA,embedded
range.webp,image+resolutions,/products,AI MODIFIED,IPTC,yes

.media { position: relative; container-type: inline-size; }
.media .ai-label {
  position: absolute; right: 2cqw; bottom: 2cqw;
  width: clamp(64px, 9cqw, 80px);
  print-color-adjust: exact;
}

exiftool -overwrite_original \
  -XMP-iptcExt:DigitalSourceType="http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia" \
  image.webp