Labelling AI-generated images.
Reviewed
Short answer
You must visibly label an AI-generated or AI-edited image in the EU when you use it professionally and it is a deepfake: it realistically resembles an existing person, object, place or event and could pass as authentic. Fantastical images (a sphinx over the Eiffel Tower) and minor edits such as colour correction usually need no label, while the provider of the AI tool must still embed a machine-readable mark in every image it generates.
Official sources: Commission guidelines on Article 50, section 6.1 · AI Act, Article 50(2) and (4)
Deepfake or not: the Commission’s image examples
| Image | Deepfake? | Why |
|---|---|---|
| AI-manipulated image of two real professional footballers in front of a building resembling a stadium | Yes | Real people, realistic scene, could pass as a real photo |
| AI image of a product in an advert or on packaging that makes it look different, better or more appealing than in real life | Yes | Misleads about the actual product |
| A real product (a car) shown against an AI-generated background, without misleading about the product | No | The AI part does not change what the product is |
| Colour correction, background extension or re-scaling of product images for aesthetic purposes | Usually not | Minor impact on how authentic the advert seems |
| Substantial AI editing of background details in a journalistic photo | Can be | Goes beyond standard editorial practice and can change how authentic the image seems |
| A sphinx flying over the Eiffel Tower | No | Unrealistic, no potential to mislead |
| An AI-generated cartoon of a historical photo | No | Clearly not a photograph |
The test is whether the image would falsely appear authentic or truthful to the audience you can reasonably expect, including children or older people if they are part of it. Photorealism makes a deepfake more likely, but the guidelines say it is not decisive on its own; the message and the context in which the image is shown also count.
Who carries which duty
| Who | Duty | Article |
|---|---|---|
| The AI image generator’s provider (for example the company behind the app) | Machine-readable mark on every image, detectable as AI; under the Code of Practice, signed metadata plus an imperceptible watermark | 50(2) |
| You, using the image for work | A visible disclosure on images that are deepfakes, at the first exposure | 50(4), 50(5) |
| You, using the image in an evidently artistic, satirical or fictional work | A disclosure that does not hamper the display or enjoyment of the work | 50(4) |
| You, for purely personal use | None under the AI Act | 2(10) |
Platform rules for images
| Platform | What it says | Source |
|---|---|---|
| Facebook, Instagram, Threads | Adds “AI info” labels when it detects industry-standard AI image indicators or when people disclose that they upload AI content; content only modified or edited with AI gets the label in the post’s menu. | Meta |
| Adds an “AI modified” label when it detects AI from IPTC metadata or its own classifiers, or when the content owner indicated AI; labels on Pins can be appealed. | Pinterest Help | |
| Shows a Content Credentials icon on images and videos signed with C2PA; clicking it shows whether AI was used and who signed. | LinkedIn Help | |
| Amazon KDP | Requires publishers to tell Amazon about AI-generated images, including cover and interior images; AI-assisted edits need not be disclosed. | KDP |
How to label an AI image
- Keep the file the AI tool gave you; many generators embed C2PA Content Credentials (which tools document it).
- Decide whether the image is a deepfake for your audience (table above). If it is, plan a visible label.
- Place the label on the image where nothing covers it; the Code of Practice gives the top right corner as an example. The EU icons are optional.
- Avoid re-saving through tools that drop metadata; screenshots and re-encoding strip it.
- Tick the platform’s AI setting when you upload, and check the published copy.
Questions
- Is an AI-generated illustration a deepfake?
- Usually not, if it is clearly not a photograph: the Commission’s guidelines give an AI-generated cartoon of a historical photo as an example that is not a deepfake. A deepfake must resemble existing subjects and falsely appear authentic or truthful to its audience.
- Do I need to label a photo I edited with AI?
- Only if the edit makes it a deepfake. The guidelines treat minor changes such as removing a passer-by, lighting adjustments, colour correction or cosmetic enhancements as usually not changing how authentic a photo seems, while substantial edits to news photos can.
- Does the label have to be on the image itself?
- The AI Act requires a clear and distinguishable disclosure at the latest at first exposure. The Code of Practice, for its signatories, says the icon or label is embedded in the content unless an equivalent alternative such as an on-screen overlay is available, and the EU icons page asks that it stays visible when content is reshared or downloaded.
Sources
- Guidelines on the transparency obligations under Article 50 of the AI Act (C(2026) 5054, 20 July 2026)European Commissionread
- AI Act, Article 50: Transparency obligationsEuropean Commission, AI Act Service Desk (AI Act Explorer)Consolidated text as at 27 July 2026 (includes the Digital Omnibus, Regulation (EU) 2026/1744), shown in the Commission’s AI Act Explorer.read
- Code of Practice on Transparency of AI-Generated Content (final text, PDF)European Commissionread
- EU Icons for labelling AI-generated contentEuropean Commissionread
- Our Approach to Labeling AI-Generated Content and Manipulated Media (updated 1 July 2024)Meta Newsroomread
- Gen AI labelsPinterest Helpread
- Content credentialsLinkedIn Helpread
- Content Guidelines — AI-generated contentAmazon Kindle Direct Publishingread
Last reviewed by Volkan Çatak, Founder, Proofcan.
Not legal advice. This page explains official texts in plain words and links each one; check the sources, or a lawyer, before relying on it for a decision.