ColophonAI content disclosure, Bristol

What a correct label looks like

The Commission’s Code of Practice sets out how an AI label must look and where it must sit, in more detail than most teams expect. Here is the label we built to it, applied correctly, and the four mistakes we find most often.

The specimens were drawn by AI, using computer code. They’re stylised line drawings that couldn’t pass as photographs, so none of them legally needs a label. Each stands for a real scenario, described underneath, and where it carries a label, the label is there to demonstrate.

The label itself

Generated
The capital letters AI, exactly the same height
The Code asks for letters of the same vertical dimension. In many typefaces the point of a capital A rises slightly above the other capitals, so that it looks the same height to the eye. Read strictly, that fails. So this mark is drawn, and the A has a flat top.
Built for any background, and for assistive technology
The white edge holds it apart from dark images and the black plate from pale ones. Screen readers, which read pages aloud for people with sight loss, announce it before the image, and it keeps its colours in Windows High Contrast Mode.
Typed. The pointed A rises past the line. Exaggerated here so you can see it.
Drawn. Both letters stop at the same line.

On an image

Tap or click a label to see its second layer, the extra detail the Code encourages.

AI-generated image. Show how it was made.

Generated by AI: the whole image.

Entry in this site’s record

Correct

Scenario: a product image made entirely by an image model. Label in the top right corner with nothing over it, visible from the first moment. Measures 1.1 and 1.2.2(a).

AI-modified image. Show how it was made.

Modified by AI: a feature the real product doesn’t have was added to a photograph.

That is the scenario. The drawing itself was generated with AI.

Entry in this site’s record

Correct

Scenario: a product photograph where AI added a feature the real product doesn’t have. The second layer, open here, says what changed. Measure 1.1(b).

Correct, with no label

Scenario: the product photographed as it is, with the setting generated. Whether a generated setting needs a label depends on what it implies, and we record the reasoning for every one. Commitment 2.

Where it goes wrong

These are the four mistakes we find most often, marked up the way a proofreader would.

Shop now

Wrong

Bottom corner, half under the shop button. Something sits between the viewer and the label, and on a phone the button covers it completely. Measure 1.2.2(a): no intervening overlay.

ai

Wrong

Faint grey “ai” in lower case. It fails contrast, the letters aren’t capitals, and a screen reader never announces it. Measure 1.1(a) and the accessibility requirements.

i

Wrong

An information button the viewer has to tap before learning anything. Recognition mustn’t depend on the viewer doing something. Measure 1.2.1(a).

Wrong

Letters stretched sideways to fill a slot in the layout. The proportions have to survive resizing. Measure 1.1(a).

Video, audio, text and venues

Every format has its own rules. Video has to bring the label back in certain circumstances, and audio needs a spoken disclaimer. Text, exhibitions and internal training material each follow rules of their own, and artistic work gets different placement again.

We set all of this out for your formats as part of an audit.

Email three links