When people hear “deepfake” they picture a fake video of a politician or a cloned voice on a scam call. The EU AI Act uses the word more broadly. A smiling AI “customer” on a bakery’s homepage can fall under it. A cartoon cat in a spacesuit almost certainly doesn’t.
This guide covers the legal definition, how the European Commission says to read it, and a pile of everyday examples so you can make the call quickly.
A tool, not legal advice. The Commission’s guidelines are not binding, and there are borderline cases. If an image or video really matters for your business, ask a lawyer.
What the law says
Article 3(60) of the AI Act defines a deep fake as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.
Both halves matter. The content has to look like something real, and it has to be the kind of thing someone could believe.
Article 50(4) then says deployers who publish deepfakes must disclose that the content is artificially generated or manipulated. That has applied since 2 August 2026.
There’s a lighter version for work that is evidently artistic, creative, satirical or fictional. You still disclose, but in a way that doesn’t ruin the piece. Think of a credit line instead of a big banner across the image.
How the Commission reads it
The Commission finalized its guidelines on Article 50 in July 2026, and a few points help a lot in practice.
The test is objective. It doesn’t matter whether you meant to fool anyone. What counts is whether the content could be taken for real.
Photorealism doesn’t decide it either. A slightly stylized image can still be a deepfake if it plausibly shows a real person or event, while a perfect render of a dragon isn’t one.
The subject has to be plausibly real. Something wholly fantastical, which nobody could reasonably believe exists, falls outside the definition.
And there’s no retroactive duty. Content generated before 2 August 2026 doesn’t have to be relabeled.
So the question to ask isn’t “how realistic is this?” It’s “could an ordinary person believe it shows something real?”
Everyday examples
Label these
AI “team photos” or “happy customers” that look like real people are the classic case. Visitors will assume those are your staff or your clients.
A realistic AI photo of your shop, office or restaurant that doesn’t match the real place also counts. It looks like a real location and could pass as a real photo.
The same goes for an AI avatar presenting a video, and for a synthetic or cloned voice of a real person in a podcast or ad.
Real photos of real events where AI added or removed people belong here too. Removing a stranger from the background of a holiday snap is one thing. Adding a crowd to a half-empty launch event is another.
Last, product photos where AI puts your real product into a realistic scene that never happened, if the result could mislead people about what the product looks like or does.
No legal duty to label
Cartoons, illustrations, icons and clearly stylized art. Abstract backgrounds, gradients and patterns. Obvious fantasy, like dragons, aliens or that cat in a spacesuit. Small edits that don’t change the meaning, such as removing dust, fixing exposure, cropping or upscaling. And obvious mockups, like a product floating in empty white space.
The gray zone
Some images sit in the middle, and these are the ones people actually argue about.
Generic AI stock photos of “a person using a laptop” don’t claim to show anyone in particular, but they do look real. A lot of site owners label them just to be safe.
AI lifestyle shots of products in realistic settings are similar. If the scene could be mistaken for a real photo shoot, a small caption is the careful choice.
Then there’s the real photo with an AI background swap. The person is real and the place isn’t. That one’s tricky.
For anything in the gray zone, a short caption like “Image edited with AI” is cheap and honest. It also ends the debate.
And text?
The deepfake rule covers images, audio and video. Text has its own rule in the same Article 50(4). AI-generated or manipulated text published to inform the public on matters of public interest has to be disclosed, unless a person with relevant competence reviewed it and someone takes editorial responsibility.
The guidelines are clear that ordinary AI fiction, product descriptions, and a chatbot reply that only the person asking ever sees aren’t “public interest” text. More on that in Do you need to label AI-written blog posts?
What a good label looks like
Put it on the content or right next to it: a caption under the image, a label on the video. It should be visible straight away, not hidden behind a click and not only in metadata. Plain words help. “AI-generated image” says more than a symbol on its own. For audio, a short spoken note at the start works.
The EU Code of Practice offers common icons for this, but they’re voluntary. See The EU icons for AI content.
A 30-second test
First, was it made or meaningfully changed by AI? If not, you’re done.
Second, does it show a real-looking person, place, object or event? If it’s clearly fantasy or abstract, you’re done.
Third, could an ordinary visitor believe it’s real? If yes, label it. If you can’t decide, label it anyway. An extra caption costs you nothing.
How AuthorLabel helps
It spots AI files. When you upload to the Media Library, AuthorLabel looks for C2PA content credentials, the IPTC Digital Source Type field (values like trainedAlgorithmicMedia) and generator names in EXIF and XMP. If it finds any, it suggests a label. Nothing changes until you save.
It shows the caption. Labeled images, video and audio get a small caption such as “AI-generated image” or “Image edited with AI”. That works in the block editor, the classic editor, Elementor image widgets and WooCommerce product galleries.
And it keeps a record. Every label change goes into the activity log with the date, the person and the reason. For gray-zone images, that’s where your decision lives.
AuthorLabel doesn’t decide for you whether something is a deepfake. You make the call. The plugin makes the label and the record easy.
In short
A deepfake is AI media that resembles something real and could pass as genuine. Intent doesn’t matter, and photorealism alone doesn’t settle it. What settles it is whether the subject is plausibly real. Label AI people, real places and realistic scenes. Cartoons and fantasy carry no duty. In the gray zone, add a short caption and move on.
Sources
- Regulation (EU) 2024/1689 (AI Act), Article 3(60) and Article 50(4) and (5).
- European Commission, Guidelines on the transparency obligations under Article 50 of the AI Act (July 2026). digital-strategy.ec.europa.eu
- Paul, Weiss, “EU finalises transparency rules for AI-generated content” (August 2026). paulweiss.com
The Commission’s guidelines aren’t legally binding.
Last reviewed: November 2026. AuthorLabel is a tool, not legal advice.