The EU's Three AI Disclosure Icons
The EU's marking Code of Practice finalised three official disclosure icons. Where they belong on a video, and why they stack with platform labels.
On 20 July 2026 the Commission finalised its guidelines on transparency obligations under Article 50, and alongside them a Code of Practice on marking AI-generated content that includes three official EU disclosure icons. Article 50 itself applied from 2 August 2026.
The question every creator asked in the fortnight after was which icon to use. That's the least interesting part. The icons are a small set with a fairly obvious mapping, and the Commission publishes them. The parts that actually cost people time are where the mark goes on a video that gets cropped, resized and re-shared, and what it does and doesn't replace once a platform has already stamped its own AI label on the same upload.
Get the assets from the source
One instruction before anything else: pull the icon files from the Commission's own materials.
Do not redraw an icon from a description, including this one. Do not recolour it to match your brand, do not restyle the stroke weight, do not substitute a similar glyph from an icon library. A disclosure mark works because it's recognisable as the mark, and an approximation is worse than no icon at all, because it signals compliance while not being the thing.
This matters more than it sounds like it should, because the icons have been described secondhand in a lot of coverage since July and the descriptions don't agree. Treat any version you didn't download from the Commission as unverified.
The icon is one layer, and not the load-bearing one
The most common mistake is treating the icon as the compliance answer. It isn't, and the structure of Article 50 makes that explicit.
| Layer | What it is | What it does not do |
|---|---|---|
| Machine-readable marking | Provenance metadata and watermarking embedded in the file, per Article 50(2) | Nothing a human viewer perceives |
| Human-visible disclosure | The icon, plus any accompanying text on the video or in its description | Nothing a machine can reliably parse from the pixels |
| Platform-native label | The "AI info" style badge a platform applies in its own UI | Doesn't travel with the file off-platform |
The Commission's own position on the first row is the one to internalise: no single technique currently meets the machine-readable marking standard, which is why a multi-layered approach combining metadata and watermarking is what's expected. Anyone who tells you C2PA credentials alone close out Article 50 is describing half a layer of a three-layer stack.
There's also a date buried in this that catches people. Systems already on the market before 2 August 2026 have until 2 December 2026 for the machine-readable marking duty under 50(2). We wrote that up separately as the December deadline nobody diarised, because it means a tool you use today may not be marking output yet and may still be inside its window.
Penalties in this regime reach €15 million or 3% of worldwide turnover, which is a number aimed at providers rather than at individual creators, but it explains why the tooling side is moving fast.
Where the icon belongs on a video
The operative standard is perceptibility: the disclosure has to be apparent to the person watching, in the format they're actually watching. That single sentence rules out most of the placements people reach for first.
Placements ranked by how well they survive real distribution:
Burned into the frame, in the safe area. Survives download, re-upload, cross-posting and screen recording. This is the only placement that still exists after someone rips your video and reposts it, which is also the scenario where disclosure matters most. Treat it like burned-in captions: if it isn't in the pixels, assume it won't make the trip.
An opening title card. Clear, unmissable, and easy to standardise across a series. The weakness is that autoplay feeds often start mid-scroll and short-form viewers skip the first beat, so a card alone can be missed on the surfaces where most views happen. Good as a component, weak as the only measure. A standard title card added to the front of the timeline is a two-minute job once you've built it.
A persistent corner bug. Small, always present, low creative cost. The catch is placement discipline, covered below.
A first-appearance lower third. Shows on screen for a few seconds early in the video, then clears. Reasonable middle ground for longer formats where a permanent bug is intrusive.
Description or caption text only. Weakest option. It's stripped by re-shares, invisible in most feed contexts, and absent entirely if the video is downloaded. Use it in addition, never instead.
The safe-area rules that break placements in practice. Vertical feeds put UI chrome over the bottom quarter of the frame and the right edge, so a bug in the bottom-right corner is frequently obscured on the exact surface that matters. Crops from 16:9 to 9:16 cut the left and right thirds, taking a corner bug with them. Autoplay-muted playback means anything spoken rather than shown doesn't count.
The workable default: top-left, inside a margin of roughly 8% from each edge, with the icon at a size that stays legible at phone scale rather than at editor scale. Then check it in the aspect ratio you're actually publishing, not the one you edited in. Adding it as a text and graphic overlay on the timeline means one edit produces every crop.
Why it stacks with the platform label instead of replacing it
Platforms already apply their own AI badges, and it's tempting to conclude that a video which got labelled automatically is covered. Three reasons that's wrong.
Different triggers. A platform's badge fires on that platform's threshold, which is usually built around realistic synthetic media and its own detection signals. A legal disclosure duty runs on a different definition and doesn't care whether the platform's classifier happened to fire.
Different surfaces. The platform badge lives in platform chrome. Your video also runs as a paid ad, in an email, on a landing page, in a client deck and inside someone else's re-upload. The badge exists in exactly one of those.
Different durability. The badge is applied at render time by the platform and disappears the moment the file leaves. A burned-in mark doesn't.
The correct mental model is that they're independent signals serving different audiences, and a mismatch between them is what looks bad. If the platform badge says AI-generated and your video carries no visible disclosure, a viewer reads that as the platform catching you rather than you telling them. Both present, saying the same thing, reads as normal.
The icon does not replace the disclosure sentence either. An icon plus a short line of plain language is the combination that actually communicates, and we've covered how to write that line so people don't scroll past it. The same logic applies when the video is also a paid placement, where the AI disclosure and the sponsorship disclosure are separate obligations that both have to land.
A working setup, once
- Download the official icon assets. Store them in your brand kit next to the logo, not in a compliance folder nobody opens.
- Build one overlay: icon plus a short disclosure line, positioned top-left inside an 8% margin, sized for phone playback.
- Add it to the timeline as a reusable element and save the edit as a draft so the next video inherits it instead of rebuilding it.
- Check placement on a 480p preview pass, which is free and carries a short per-user cooldown, before the final export. The preview and export split exists exactly for this kind of check, since the export itself is charged once regardless of how many clips are on the timeline.
- Verify in every aspect ratio you publish, not just the master.
- Keep using the platform's own disclosure toggle. It's a separate signal and it costs nothing to be consistent.
FAQ
Which of the three icons applies to a normal AI-generated marketing video?
The mapping is set out in the Code of Practice itself, and it's worth reading there rather than taking a summary, because the distinction between the icons turns on the kind of AI involvement rather than on the content type. Pulling the mapping and the artwork from the same source is also the only way to be sure the icon you ship matches the case you're claiming.
Does an icon on the video satisfy Article 50 by itself?
No. Visible disclosure and machine-readable marking are separate layers. The Commission has said no single technique currently meets the machine-readable standard, so metadata and watermarking are expected together, and neither of those is something a viewer sees. An icon closes the human-visible layer only.
What if the platform already labelled my upload?
Keep your own mark anyway. The platform badge is applied in platform UI, on that platform's trigger conditions, and it doesn't survive the file being downloaded, cross-posted or embedded elsewhere. They're independent signals and they should agree with each other.
I'm outside the EU but some of my audience isn't. Does this apply?
The relevant question is where the content is distributed and to whom, not where you sit. If you're deliberately reaching EU audiences, the obligations are in scope regardless of your own location. A single burned-in disclosure applied to everything is usually cheaper than maintaining two versions of a video, and it's rarely a real creative cost once the overlay is built.