EU icons: generated vs modified

Pick the EU icon by how much of the file is synthetic. Fully AI-generated when the entire deepfake (or public-interest text) is generated, with no human-created elements beyond prompting. Partially AI-modified when a real plate was altered into a deepfake. The third mark is the basic icon, usually with words or an interactive second layer. Download the Commission files. Do not redraw them.

Article 50 has applied since 2 August 2026. Guidelines 20 July 2026. Fines can reach €15 million or 3%. The icons live on the Commission's page: EU icons for labelling AI-generated content. Using the icons is optional. Labelling is not. The icon does not, by itself, establish legal compliance.

This page is generated versus modified. The interactive second layer is EU AI icons have a second layer. The paste lines are the disclosure template.

Three official marks, one truth per file

The Commission published three icons, each in four visual variations: black, white, black at 50 percent transparency, white at 50 percent transparency. SVG and PNG zips are on that page. Signatories of the Code of Practice have placement specs. Non-signatories may still use the files. Using them is not a claim that you signed the code.

Fully AI-generated. The entire deepfake content (image, audio, video) or text is fully generated by AI with no human-created elements or human editorial control apart from prompting. Commission examples: fully AI-generated deepfake videos with politicians or fictional events; fully AI-composed music or art (which may also sit in the artistic limb); AI-generated news summaries. A talking head that was never in a room. A Veo take of a street that was never filmed. If that is this week's file, this is the icon, plus a line like "This video is AI-generated."

Partially AI-modified. Pre-existing, human-made content was partially modified with AI, turning it into a deepfake or into public-interest text. Commission examples: a face in an authentic photograph swapped for a politician; authentic photographs of an empty apartment furnished with AI. You shot the room; a model restyled the sofa. You filmed a real presenter; you swapped the mouth line they did not say. You are here, not on fully-generated.

Basic icon. When AI was involved in creating a deepfake or published text, especially when you pair the mark with a custom text label or an interactive second layer. Commission example: a deepfake video with the text "voices generated with" followed by the basic icon. User-testing improved when the basic icon sat with words. Pair it. Do not ship the mark alone and call it done.

The fork is not "which icon is quietest." It is what is true of the bytes. A fully generated person composited over a real product still plate is a mixed file; the talking-head layer is generated, the pack may be photographed. If the deepfake the viewer would take as authentic is the person, you still label. Prefer the icon that matches the synthetic part you are disclosing, and put plain words next to it. When the whole picture is a generate, use fully-generated. When a camera original was altered into the misleading resemblance, use modified.

How to pick without a philosophy seminar

Ask two questions in order.

Would an EU viewer take this as a real person, place, or event? If no — clearly a cartoon, clearly a game, clearly not a person — you may still disclose as house style, but you may be outside the Article 50 deepfake radio. If yes, continue.

Was the whole picture generated, or was a real recording modified? Whole picture generated → fully-generated icon and line. Real recording modified → modified icon and line. Real picture, synthetic voice → basic icon plus "The voice in this video is AI-generated" is the shape the Commission already used.

Standard assistive editing is exempt: colour, crop, denoise, captions on a real recording. A colour-graded interview of a real person is not a deepfake because you used a timeline. Do not put a modified icon on a grade and train the audience to ignore icons.

Do not invent a fourth icon. Do not ship a custom SVG and call it the EU mark. Do not treat a TikTok sparkle or a YouTube overlay as the Commission file. Platform badges are contract. These three files are the optional official marks for synthetic media disclosure.

First exposure still rules placement. The Commission summary: the icon should be clearly perceivable and distinguishable at the latest at first exposure; placed where no intervening overlay elements exist; directly embedded into the deepfake or published text (except for creative works), unless an equivalent user-interface overlay is available; visible when content is reshared or downloaded. If the disclosure appears for a limited time, it should remain long enough to be read.

You are usually the deployer of the file you publish. Provider marking (machine-readable) is a different artefact, with grace to 2 December 2026 for systems already on the market before 2 August 2026. A provenance tick is not your on-screen icon.

Mixed files and the honest radio

Production is rarely a pure generate or a pure grade. Typical mixes:

• Photoreal presenter, real pack on the table. The person is the deepfake-shaped part. Label the person. Fully-generated is the honest radio if the person never existed. Do not hide behind the pack photo.
• Real presenter, generated background or extra. Modified, if the result would still appear as an authentic scene. A clearly impossible backdrop is a weaker resemblance case; a generated "real" street is not.
• Real cut, lipsync pass that puts new words in a real mouth. Modified. Making a real person appear to say something they did not is the classic deepfake.
• Real cut, TTS bed, no mouth change. Voice is the synthetic part. Basic icon plus a voice line is the Commission's own example shape.
• Slideshow of real product stills, AI script, burned captions. Usually not an Article 50 deepfake because of the stills and the assistive text. Platform rules may still want nothing. Do not slap fully-generated on a photo gallery because a model wrote the hook.

When two icons could argue, pick the one that describes the misleading layer, and write the words so a person does not have to guess. Words beat a silent mark. The Commission said user-testing improved when the basic icon had a text label.

Artistic work may label without spoiling the work. Ads and UGC testimonials are not that quieter path. Article 50 artistic exception is narrow.

Lock the pick before upload

Do the radio in the edit, not in the caption box after the media buy.

1. Name the file's true state in the brief: generated, modified, voice-only, or assistive-only.
2. Download the Commission PNG or SVG. Place it where chrome will not cover it.
3. Burn the matching paste line. Keep it up long enough to read.
4. Then set YouTube, TikTok, and Meta toggles. The AI content label is extra.

A UGC video generator will not pick the icon. Veo 3.1 will not either. If you want the mark on the same timeline as the take, sign up and treat the icon pick as picture lock.

Content generated and published before 2 August 2026 is not retroactive. Generated before and published after is covered. Personal, non-professional use sits outside the Act. If you are reading this as a checklist, assume you are the deployer.

FAQ

Do I have to use the EU icons?

No. Labelling covered content is required; the icons are optional. If you use an icon, use the Commission files. User-testing favoured icon plus plain words.

What if half the frame is generated and half is a real product?

Disclose the synthetic part that a viewer would take as authentic. A real pack does not launder a fake person. Fully-generated is usually honest for a person who was never filmed; say so in words.

Can I use fully-generated on everything to be safe?

Over-labelling a real, unaltered recording is a different kind of false. Pick the true radio. Assistive editing of a real interview is not "fully AI-generated." Save the loudest mark for the files that earned it.