Placing the AI label: where the EU icon belongs in an image

Last updated: August 19, 2026

The obligation is settled, the execution is not. You know your AI generated creative needs a visible marking. What stays open is the question that shows up in the layout. Where does it go? Bottom right on every image, whatever happens to be there? As small as possible so the ad does not suffer? And what do you do when it is not one image but three hundred.

This article answers both: what the law actually asks of placement, and which rules hold up in practice. It ends with the way to stop deciding this by hand.

In short

• No pixel rules in the law. Article 50 of the EU AI Act prescribes neither a corner nor a size. It requires the disclosure to be clear and distinguishable, at the latest when the content is first perceived.
• "Clear and distinguishable" is a standard, not a free pass. A label that disappears in a crop, sits on the headline or is unreadable at five percent of the image width does not meet it.
• A corner is the right answer. It is the convention for legal notices, it disturbs the creative least, and it is where viewers look for it.
• Contrast decides visibility. The official EU icons come in black and white for exactly this reason.
• Consistency matters across a series. A label that lands somewhere else on every asset looks bolted on afterwards and makes a later review harder than it needs to be.
• From August 2, 2026 the visible marking under paragraph 4 applies directly.

What the law says about placement

Article 50(4) of the AI Act obliges you as a deployer to disclose that content has been artificially generated or manipulated. The text names a requirement for the manner: the disclosure has to be clear and distinguishable, at the latest at the time of the first interaction or exposure.

What it does not say matters just as much: no minimum size in pixels, no prescribed corner, no particular wording. The regulation describes a goal, not a layout. The icons the European Commission provides for labeling AI content are an offer that makes the goal reachable, and they are free to use.

For you that means design freedom with the burden of proof attached. In a dispute it does not count that an icon was somewhere, it counts whether an average viewer perceives it without effort. Measure your placement against that.

Five rules that hold up in practice

1. Take a corner, preferably a lower one

Legal notices live in corners and viewers look for them there. In nearly every piece of advertising the top carries the headline and the logo, while the bottom more often has open space. A corner disturbs the composition least, because the main subject is rarely there.

The border is the second choice: when all four corners are occupied, move along the edge to the next quiet spot instead of sliding into the image. A label in the middle reads like a watermark and looks like a mistake.

2. Keep clear of anything that carries meaning

Face, product edge, logo, price, headline: wherever the label lands on one of those, the ad loses and the marking loses with it. An icon sitting half on a bold headline is harder to read than one three percent further down in empty background.

Plan for the crop as well. What runs square in the timeline loses space at the top and bottom in a 9:16 cut. A label right at the border can vanish exactly there. A few percent of air to the edge solves it.

3. Pick the variant by the background, not by habit

The official EU icons come in a black and a white variant. That is not a design detail, it is the reason the marking is as readable on a dark photograph as on white. Black on dark water is not a marking, it is a smudge.

Judge the spot the icon sits on, not the image as a whole. A mostly bright creative can have a dark shadow in exactly the corner you picked.

4. Small is good, too small is useless

There is no legal minimum size, but there is a practical one: the icon has to stay recognisable on the device the content is consumed on, and that device is a phone. In our work with advertising creatives a corridor of roughly 8 to 13 percent of the image width has proven usable. Below that it turns into graphics without a message in a feed format, above it starts to dominate the creative.

Opacity is the second lever. The EU also provides variants at 50 percent transparency. They are more discreet, which is exactly why they belong on a calm, high contrast background only.

5. Keep it the same across the campaign

When thirty ads of one campaign carry thirty different positions, that does not look like a process, it looks like handwork under time pressure. For the brand it is restless, for a later proof it is needlessly complicated. Decide the placement once, keep it in a template, and deviate only when a creative forces it.

The four most common mistakes

The label sits in the crop. A classic when a creative runs in several aspect ratios. Check the marking in every format that actually goes live, not only in the master.

The label sits on the headline. This happens whenever a fixed position meets a layout that did not plan for it. Exactly here it pays to look per creative rather than per campaign.

The wrong variant. Black on dark, white on bright. The most common reason: the position was set once and the colour never touched again.

Only marked machine readably. The XMP/IPTC marking matters, but it covers the provider duty under paragraph 2. The visible marking you owe as a deployer under paragraph 4 is not replaced by it, all the more since platforms routinely strip metadata on upload.

When it is not one image but three hundred

For a single creative, placement is a minute of work. For a product catalogue, an ad batch or a running creative pipeline it is a process. Two ways work.

Templates. You set position, size, variant and text label once and apply them to every job. That is the right way for series that look alike, packshots on the same background for instance.

Automatic placement. Where creatives differ strongly, a fixed position is bound to be wrong sometimes. That is what we built automatic placement for. A vision model looks at every image before it is rendered and decides three things: where the label sits, how large it is and whether the light or the dark variant carries.

The rules from this article are guard rails there, not advice. The model is steered towards the corners and leaves them only when all four are occupied. The point is held in the band along the border, so a decision can never land in the middle of a creative. The size is capped at both ends, so the marking stays recognisable without dominating the image. And the colour variant follows the brightness of the exact spot that was chosen. What the model may not do: add a label, pick a different icon or change anything else in your settings. If it returns nothing usable, the placement you sent along simply stands.

The feature runs through the API, is marked as beta and is part of Enterprise. It uses your own model account, so the analysis is billed there. The technical details are in the API documentation.

Frequently asked questions

Does the EU AI Act prescribe a position for the AI label?

No. Article 50 requires a clear and distinguishable disclosure at the latest when the content is first perceived, but it names neither a corner nor a size nor a wording. That freedom comes with the burden of explaining, if it ever comes to that, why your implementation was clear and distinguishable.

How large does the EU icon have to be?

There is no legal minimum. In practice a corridor of roughly 8 to 13 percent of the image width works, because the icon has to stay recognisable on a phone. What counts is not the percentage but whether a viewer perceives it without effort.

May I use the transparent variant of the icon?

Yes, the EU provides variants at 50 percent transparency. Use them on a calm background with clear contrast. On a busy photograph a half transparent icon loses so much legibility that the marking misses its purpose.

Does the label have to be on every format of a campaign?

Yes. Every creative that goes live is content of its own and has to carry the marking. Watch the crops in particular: a label that sits perfectly in the 1:1 master can be cut off in the 9:16 version.

What happens in automatic placement when the model gets it wrong?

It can only get it wrong within fixed limits. The point stays in the band along the border, the size stays in the permitted corridor, and the icon stays the one you chose. If no usable answer comes back, the placement from your request is used and the job does not fail.

Does automatic placement replace a human check?

No. It takes the per creative decision off your hands, not the responsibility. For a campaign with a lot riding on it, look at the finished assets, the way you would with any other automated step in production.

What now?

If you label individual images, a corner, good contrast and a look at the result is all it takes. That works directly in the tool, free and without an account. If you produce in series, fix the placement in a template and let automatic placement decide where the creatives differ.

Read on: The complete guide to AI labeling · Labeling AI ads on Meta · The official EU icons at a glance