Your AI label has to move. Your brand guidelines cannot tell it where.
Executive Summary
- The binding law says disclose, and says nothing about where. Every placement rule comes from the EU Code of Practice, which is voluntary: the label sits where no other overlay element already is, holding clear spacing from logos, subtitles and on-screen text. No single agreed position satisfies that across markets and formats.
- Localisation changes the geometry, not only the words. A German headline runs to two lines, a longer currency string pushes the price flash down, a market mandates its own badge in a fixed position. The space left free for the label in one market's asset is not the space left free in another's.
- The label may sit on the picture itself, over a face or a product, but never on a logo, a subtitle or on-screen text. In a video it has to keep reappearing, and each appearance meets a different frame.
- The number of placement decisions equals the number of market, format and brand line combinations you ship, and nobody is auditing those one by one.
- Nothing in a DAM or a PIM records which part of a composition came from a generative module, where it sits in the frame, or at which second it appears. A yes or no flag on the file is as far as most schemas go.
This is not legal advice. It is a reading of the Act, the Commission's guidelines and the Code of Practice as they stand in September 2026, and the last two can still change.
Why one fixed position for the AI label does not work
Right now, in rooms across your organisation, somebody is deciding where the label should go. The answer coming back might be a corner, usually the top right, picked once off a single reference layout. That corner is not a fixed quantity.
The top right of a 16:9 film and the top right of a 9:16 vertical are different pieces of space, so a position that worked in one crops badly in the other. Then the corner fills up. A headline runs to two lines in German where one line did the job in English. A longer currency string pushes the price flash down. A legal line takes three lines in one market and one in another. A local retailer's logo arrives at an aspect ratio you did not design around, and where a market mandates its own badge in a fixed position, it reserves that space before your label gets near it. Every one of those redraws the canvas, so the space left free in the German asset is not the space left free in the Spanish one.
You can see all of that while you build the asset, which is why the working answer today is a designer moving the label around each frame until it looks acceptable, one asset at a time. The channel logo a broadcaster burns in after you ship, you cannot see at all. The AI label turns up last, into a layout already built around everything else.
The Commission published its guidelines on the Article 50 transparency obligations on 20 July 2026, alongside a Code of Practice carrying the design, placement and timing detail. Read it as an operator rather than a lawyer and it comes down to one property: what else is in the frame decides where the label sits, and in a video the frame keeps changing. It changes with the format, with the market, with the second of the runtime, and with the brand line.
What the EU AI Act says about where the AI label goes
An AI disclosure label is the visible mark that tells a viewer content was generated or manipulated by AI. The Act itself requires disclosure that is clear and distinguishable by the time somebody is first exposed to the content, and says nothing about position. The position rules are all in the Code of Practice, which is voluntary. Non-signatories still have to reach the same place by other means, and the Commission's guidance tells them to run a gap analysis against the Code to show they have.
The Code is specific about where it goes:
"The icon or equivalent label will appear in an appropriate place where no intervening overlay elements exist (e.g., in the top right corner of an image or video deep fake)." - EU Code of Practice on Transparency of AI-Generated Content (2026)
It names the top right corner and in the same breath forbids the label from sitting where an overlay element already is. In broadcast that corner holds the channel logo. On social it holds the platform's own badge.
A second passage goes further. The label has to be "clearly perceivable and distinguishable at the latest at the time of first exposure," achieved by "maintaining sufficient spacing to other overlay elements, disclaimers, and sound/visual on-screen/on-display elements" while remaining "visible against any background."
So the requirement is a relationship between the label and everything else in the frame.
In a video that relationship has to be solved more than once. A subtitle at second four and a sustainability badge at second twenty are different frames, so a corner that was free at the start of the spot can be occupied when the label comes back. The label moves inside one video as well as between versions of it.
A brand guideline cannot describe a relationship to something it has never seen.
Can the AI label sit on top of a person, a product or the background?
Yes. The same sentence that demands spacing from logos and subtitles ends by requiring the label to "remain visible against any background." Background is the Code's word for whatever sits behind the label, and staying visible is the whole test. A face, a table, a product, a generated sky: none of them are out of bounds. That is why the Commission ships each icon in black, white, black transparent and white transparent, so you can pick the one that survives whatever it lands on. Once the logos and subtitles have taken the corners, the image itself is the only space left to put the label on.
The Act's own words are clear and distinguishable, and the Commission reads distinguishable as "easy to identify as separate from other information and the environment in which the content is presented," failing wherever the information "can be easily overlooked or missed by natural persons under normal exposure or interaction conditions." Put a white label on the plain white background of an e-commerce packshot and it technically exists, but nobody sees it.
The label belongs on the picture itself. The Code says it "will be directly embedded into the content," and the only alternative it allows is an interface element that still looks to the viewer like it sits on the content. A caption underneath does not count.
How often does the AI label have to appear in a video ad?
"Signatories will display the icon or equivalent label at the beginning of the video as well as, where possible, at regular intervals throughout the video and, at a minimum, after interruptions (e.g., after commercial or advertising breaks)." - EU Code of Practice on Transparency of AI-Generated Content (2026)
Three requirements. The label appears at the start, always. It keeps reappearing at intervals through the runtime wherever you can manage that. And it returns after a break in playback.
Take the third off the table. That requirement lands on long-form video you own, not on a 30 second marketing spot.
If only part of the spot is AI, the Act and the Code give different answers. The Act's deadline is the moment a viewer is first exposed to the AI part, and the Commission's guidance says a deployer may disclose earlier than that but does not have to. The Code asks for the label at the beginning of the video and through it, whether the AI arrives at second two or second twenty.
The second requirement reaches every ad you make. A label that flashes once at the head of a thirty second spot has not appeared at "regular intervals throughout the video", and whatever interval you settle on, something has to stamp the label repeatedly onto a timeline.
You already cut clips. Every fifteen and six second version pulled out of a thirty second master is one, and if the label only lived in the master's opening second, every version cut from the back half ships with no disclosure on it. That is exactly what the repetition is for: the Code asks for it "to account for downstream use (e.g., screenshots and clipped fragments)."
How much design freedom do you have on the AI label?
The Code permits the label to be "expressed in different styles (e.g., contrast ratio, colour, or typography), as long as it remains clear, accessible, and distinguishable." The Commission's icon set is optional: its own page states that "the use of these EU icons is optional, but the labelling requirements under Article 50 AI Act are not."
Most teams are about to leave the upside on the table. A mandatory element you are allowed to typeset is an element that can belong to your brand: the luxury line carrying a label set in its own face and weight, the value line carrying a louder one, neither looking like a compliance sticker pressed onto the corner of a finished ad. The default, if nobody decides otherwise, is a foreign object in every frame you ship.
What the freedom costs is consistency. Every combination of market, format and brand line becomes a decision somebody has to make, each one needing a position that works in that frame, and nobody is checking thousands of renders manually.
Why partial AI labelling is hard to implement
"Signatories are furthermore encouraged to display the icon or equivalent label throughout the deep fake video or during the deep fake part of the video, where the deep fake segment does not necessarily span the entire duration of the video." - EU Code of Practice on Transparency of AI-Generated Content (2026)
You can keep the label up continuously through just the AI seconds. The Code encourages that on top of its own timing requirement rather than as a substitute for it, so under the Code the label still has to appear at the start and again across the video. It is still worth wanting: a label on the whole film tells the viewer nothing about how little of it was made by a model, so a film that is one tenth synthetic takes the same hit with viewers as one that is entirely synthetic.
Almost nothing in production can take that up. Labelling the AI part of a video means knowing the in-point and out-point of the synthetic segment, and in an edited film those boundaries are cuts, because a generated shot is a shot: it starts at one cut and ends at the next. A system that assembled the edit holds every one of those timecodes, because it made them. A finished MP4 carries none of them.
The still image version is worse, because there is no cut to fall back on. Labelling only the generated region of a photograph would mean storing that region as a region, which no rule currently requires.
The same missing record is what stops the placement rule working. A DAM stores a file and its attributes. A PIM stores product data. Neither has a field for which part of a composition came from a generative module, where it sits in the frame, or at which second it appears.
Where the AI disclosure decision has to live
Put it together.
- A position that has to clear whatever else is on screen and stay readable over whatever sits behind it.
- A label that reappears through the runtime, landing in a different frame each time.
- A treatment that changes with the brand line.
Underneath all three, a record precise enough to say where in the frame and when in the runtime the disclosed thing happened.
A guideline cannot settle any of it, because every one of those constraints depends on a frame nobody has built yet. The frame exists at one moment: assembly. What ends up in it depends on the creative call as much as the compliance one, on what a director decided should sit where. Most enterprises are still handing this to a creative team to solve while they design the ads, or worse, treating it as something to add once the ad is finished.
A system that assembles compositions already holds the record, because it could not have assembled anything without it. To put a generated background behind a real product shot, something had to know which module produced the background, that the product came from the DAM, where each sits in the frame, and at which second each appears. Where that same system called the model, it knows which layer is synthetic because it made it. That is the minimum information required to build the asset, and it exists whether or not anybody ever asks to see it.
Once the record exists, the label becomes a dynamic output the system computes rather than a fixed coordinate somebody wrote down once. We run a rules engine that does this today: it reads the position and extent of everything already placed in the composition and puts the disclosure into space that is genuinely free and pre-approved by the brand, per market and per format. Channel logo top left, the label moves to the opposite quadrant. Subtitle track present, it shifts clear of the track. No designer resets a position value for any combination. One CPG e-commerce programme we rebuilt as an engine needed 15,000 assets across 38 products: an estimated 7,500 hours by hand, and under 94 hours once the engine ran it.
That is a content supply chain built properly: one that knows every element in every composition, tracks it at the moment of assembly, and renders the correct version of everything downstream, disclosure included, for every market and every format. We build it. It is called VARYCON. If you want to know whether your stack could answer the where and when question for a single asset, that is a short conversation, and a useful one to have before a regulator asks first.
Related Questions
Where does the EU AI Act say the AI disclosure label has to be placed?
The Act requires disclosure without specifying a position. The detail sits in the EU Code of Practice on Transparency of AI-Generated Content, a voluntary instrument that non-signatories are still expected to measure themselves against, which says the label goes "in an appropriate place where no intervening overlay elements exist" and gives the top right corner as an example. A separate passage requires sufficient spacing from other overlays, disclaimers and on-screen elements, and legibility against any background. The requirement is relational, so one fixed position cannot satisfy it across markets and formats.
Can the AI label be placed on top of a person or a product in the video?
Yes. The Code of practice restricts the label from colliding with overlay elements such as logos, subtitles and on-screen text, but it treats the picture underneath as a legibility question rather than a keep-out zone, requiring only that the label "remain visible against any background." A face, a product or a table is therefore permitted. The practical limit comes from the Act's word distinguishable, which the Commission reads as "easy to identify as separate from other information and the environment in which the content is presented," so a label that blends into a plain background or a busy surface fails.
How often does an AI label have to appear in a video ad?
The Code of Practice specifies the label at the beginning of the video, at regular intervals through it where possible, and at a minimum after interruptions such as commercial breaks. The interruption requirement applies to long-form video that can be interrupted, not to a thirty second spot, which is itself the interruption. The interval requirement, which the Code qualifies with where possible, applies to every ad, and the stated reason is downstream use: screenshots and clipped fragments, including your own cutdowns, have to carry the disclosure too. Where only part of a video is AI, the Act's own deadline is the moment a viewer is first exposed to that part, while the Code asks for the label from the beginning of the video regardless.
Does the AI disclosure label have to be translated for each market?
No. The EU Code of Practice specifies the capitalised "AI" in English as the main visual element of the label, in every market. Spain, Germany, Portugal and Italy all have their own acronym for artificial intelligence, and none of that changes the label. The Code writes in one exception, and it is permissive rather than mandatory: the acronym may be disclosed in the national language where use of English is incompatible with applicable national laws on the use of languages in commercial or administrative matters. France is the market most likely to trigger it, because the Toubon Law of 1994 makes French mandatory in all advertising aimed at the French public, though whether a two-letter acronym inside an icon is caught by that is a question for local counsel. Budget for one label in English, plus a legal check in France.
Does adding an AI label hurt advertising performance?
On the asset carrying it, yes. The Nuremberg Institute for Market Decisions showed people the same advertisement and changed only whether it was described as a photograph or as AI-generated. Across samples of 1,000 respondents each in the United States, the United Kingdom and Germany, consumers "tended to see these ads as less natural and less useful, even though the content was identical to the ads labeled as human-made," and they were less inclined to click through. The effect does not appear to travel past the labelled asset: the London School of Economics, on a nationally representative sample of 3,861 people, found the aversion "limited in scope." A blanket label also communicates nothing about how little of the asset is actually synthetic.
