Do you need to label AI images on your website?

AI images are everywhere now. Blog headers, Instagram graphics, product shots with a fresh background, team photos touched up by some app. So does every one of them need an “AI” sticker? Nope. The EU AI Act only asks you to disclose deepfakes, and that’s a much smaller group than most people think.

Below we walk through how to tell them apart, using the examples the European Commission itself gives, and where to put a label when you do need one.

A tool, not legal advice. This guide explains public documents in plain language. Whether a specific image needs a label depends on context. If it matters for your business, check with a lawyer.

Two rules, two audiences

Article 50 has two rules about AI images, and they’re aimed at different people:

• Article 50(2), machine-readable marking. The tool that makes the image (the provider) has to embed a technical marker like metadata or a watermark. It covers nearly every AI image, but it’s the tool’s job, not yours.
• Article 50(4), visible disclosure of deepfakes. If you publish an AI-generated or edited image that counts as a deepfake, you (the deployer) have to say it’s artificial. This is the one that matters for site owners.

So forget “Is this AI?” The real question is “Is this a deepfake?”

What turns an image into a deepfake

The AI Act’s definition is a mouthful: 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.

The Commission breaks that down into four tests. An image has to pass all four.

1. Resemblance. It clearly looks like its subject. Doesn’t have to be a perfect copy.
2. Existing. The subject is real or could plausibly be real. A realistic person who never existed still counts. A dragon, a flying human or an elephant driving a car doesn’t.
3. The right kind of subject. People, products, buildings, places, animals, realistic scenes.
4. Looks authentic. A reasonable viewer could take it for the real thing.

That last test usually settles it. A photo-style image can still be fine if, in context, it’s obviously not real.

Deepfake or not? Real examples

The guidelines give a bunch of examples. Here are the ones that come up most on websites.

Probably a deepfake, so disclose it:

• A realistic AI image of two real, recognizable athletes outside a stadium.
• An AI video of someone who looks like a politician giving a speech.
• An ad with an AI version of a well-known influencer.
• A lifelike AI avatar of your CEO thanking the team.
• An AI product image that could mislead people about how the product really looks, works or what it does. For example, making it look better made than it is.

Probably not a deepfake:

• A sphinx flying over the Eiffel Tower.
• Mice arguing about cheese in a cheese brand’s ad.
• An AI cartoon based on a historical photo.
• Made-up forests and castles in a video game.
• A real product on an AI-generated background, as long as the product itself isn’t misrepresented.

The guidelines also list everyday edits that usually don’t make something a deepfake: color correction, lighting tweaks, noise reduction, removing random passersby, extending or swapping a background for looks, arranging existing products together, and resizing. Heavy edits that change the meaning are another story, like changing someone’s body shape or skin color, or showing events that never happened.

A quick gut check

Ask yourself three things:

1. Does it look like a real photo of a real or realistic person, product, place or event? If not, you’re probably fine.
2. Could a visitor believe it’s an actual photo? Think about who really visits your site. If your audience includes kids or older people, the bar for “could believe” is lower.
3. Would believing it change how they see the subject? A misleading product image or a fake photo of a real person matters. A nicer background usually doesn’t.

Three yeses? Label it.

Art, satire and fiction

Deepfakes that are part of an obviously artistic, creative, satirical or fictional work still need a disclosure, just a lighter one. The guidelines say it should be appropriate and shouldn’t get in the way of enjoying the work. A short credit line or a note at the top of a gallery often does the job.

Where the label goes

The Act wants the disclosure to be clear and easy to notice, no later than when someone first sees the image.

The voluntary Code of Practice gets more specific. It says to put the label somewhere nothing else is layered on top of it, and gives the top-right corner of an image as the example. It also says people should notice it without having to click, hover or squint.

So in practice:

• Good: a small “AI” badge in the corner of the image, or a short caption right under it, like “AI-generated image”.
• Bad: a note in the footer, a line at the end of the post, or text that only shows on hover. By then it’s too late, or nobody sees it at all.

The EU “AI” icon

The Code of Practice also introduces a shared EU icon built around the letters “AI”. It comes in two versions: “AI + GENERATED” for fully generated content and “AI + MODIFIED” for content that’s been partly changed. The Code says the icons are publicly available to use without attribution. You don’t have to use them, though. A plain-text label like “AI-generated image” works just as well.

Old images

According to the guidelines, deepfakes generated before 2 August 2026 don’t need labels after the fact, although adding them is encouraged. New images, and old ones you reuse in new content, are a different deal. Label those as you publish.

Don’t strip the hidden marker

Lots of AI tools now tuck a C2PA content credential or an IPTC “digital source type” field into the image file. Platforms and browsers use those to spot AI content. The trouble is, some image optimization and compression plugins strip metadata on upload. If you use one, check its settings. Keeping the marker costs you nothing.

How AuthorLabel handles images

AuthorLabel deals with images in three steps.

1. It reads what the file already says. When you upload an image, the plugin checks the C2PA, IPTC Digital Source Type and generator EXIF/XMP fields. If the file says an AI tool made it, AuthorLabel suggests the right label. It never guesses from how the picture looks.
2. You make the call. Every file in the media library gets an “AI transparency” panel. Mark it AI-generated, AI-assisted or not AI, and note which tool you used. Nothing goes public without your say-so.
3. It shows the label on the page. When the image appears in a post or product, AuthorLabel adds a short caption right under it, plus machine-readable markers (Schema.org JSON-LD and a data-ai-generated attribute). Your post content and the image file stay untouched.

Got a pile of AI images from before you installed the plugin? Bulk labeling lets you filter by date, author or tool and label up to 200 items at once. Every bulk action can be undone, and the action log shows who labeled what and when.

AuthorLabel Pro adds a separate deepfake label type, handy for teams that need to keep deepfakes apart from other AI media.

The short version

• Label images that are realistic, look like something real, and could mislead.
• Stylized art, obvious fantasy and real products on AI backgrounds that don’t mislead usually don’t need a label.
• Put the label on the image or right next to it, visible without any clicking.
• Leave the AI tool’s hidden marker alone.

Sources

• Regulation (EU) 2024/1689 (AI Act), Article 3(60) and Article 50(2), (4) and (5).
• European Commission, Guidelines on the transparency obligations under Article 50 of the AI Act, C(2026) 7702 final, Section 6.1. digital-strategy.ec.europa.eu
• Code of Practice on Transparency of AI-Generated Content, deployer section, Measure 1.2. digital-strategy.ec.europa.eu

The Commission’s guidelines aren’t legally binding, and the Code of Practice is voluntary.

Last reviewed: November 2026. AuthorLabel is a tool, not legal advice.