Labeling AI images in a legally sound way (EU AI Act Art. 50) – the practical check for WordPress
Which AI images must be labeled, and which don't? Metadata vs. visible obligation, exemptions and implementation in WordPress.
Article 50 of the EU AI Act has applied since 2 August 2026. The machine-readable marking of AI-generated images is the obligation of the AI providers. As a website operator you must label deepfakes visibly and make sure that an existing machine-readable marking is not lost on your site. Not subject to labeling are images where AI only assisted, obviously artificial or clearly artistic content, and purely private use. In WordPress you cover this if a plugin sets the marking automatically on upload, writes it into all image sizes and adds a visible label that can be adjusted per image.
Which images are affected
Relevant are images that were generated or substantially altered by an AI system and made publicly available – generated content from Midjourney, DALL·E, Firefly and similar tools, as well as AI edits that substantially change the image content (generative fill, replacing parts of an image).
Machine-readable (para. 2) and visible (para. 4)
The machine-readable marking – via C2PA metadata or a watermark – is the base level for every AI image and is set by the AI provider. On your site it is about not losing it. The visible disclosure is added when the image qualifies as a deepfake: a realistic depiction of real or realistic-looking persons, places or events that did not happen this way. For clearly recognizable art or satire, a discreet disclosure is enough.
The exceptions
- AI only assisted (cropping, exposure, retouching, compression) and did not substantially change the input data.
- The image is obviously artificial and has no potential to deceive.
- Purely private, non-public use.
- Legally permitted purposes, for example law enforcement.
Implementation in WordPress
In practice this means: a plugin that detects on upload whether an image is AI-generated, transfers an existing machine-readable marking into every image size and sets a visible label that can be adjusted manually per image. A history log records which image was classified when and how.
Penalties and deadlines
Breaches of the transparency obligations of Article 50 can be fined under Article 99 with up to €15 million or 3 percent of worldwide annual turnover – for companies the higher amount applies, for small and medium-sized enterprises the lower one. Images generated before 2 August 2026 do not have to be labeled retroactively; the European Commission recommends voluntary retroactive labeling where possible.
Decision table
The machine-readable marker (para. 2) comes from the AI provider – modern tools include it. The operator’s own legal obligation is the visible labeling of deepfakes (para. 4).
| Image | Machine-readable marker | Visible label (operator’s duty) |
|---|---|---|
| Photorealistic AI image of a real person or scene (deepfake) | yes (from the AI tool) | required |
| AI-generated, clearly artificial or illustrative content | yes (from the AI tool) | not mandatory |
| Photo with AI retouching or standard editing | – | none |
| Clearly recognizable art or satire | yes (from the AI tool) | discreet, non-intrusive |
| Purely private, unpublished use | – | none |
For how to keep the machine-readable marking intact across every image size, see the guide Labeling AI images in WordPress: C2PA, thumbnails and the EU AI Act.
Disclaimer
Note: This article is general guidance and not legal advice. We accept no liability for the accuracy, completeness or timeliness of the information and recommend checking the sources yourself – the original text of Article 50 of the EU AI Act (Regulation (EU) 2024/1689) and the European Commission’s guidelines. Information on IMG Performer: img-performer.enfants.de. Regulation text: eur-lex.europa.eu/eli/reg/2024/1689/oj · Explanation: artificialintelligenceact.eu/article/50.
Frequently asked questions
Do I have to label old AI images retroactively?
No. For images the date of creation is decisive: AI images created before 2 August 2026 are exempt. Voluntary retroactive labeling is recommended.
Does AI upscaling or AI retouching count as subject to labeling?
Usually not, as long as the input data is not substantially changed. Standard editing falls under the exception.
Who is liable – the agency or the website operator?
The visible disclosure obligation under paragraph 4 applies to the operator who publishes the image. Implementation can be arranged contractually with an agency; the legal responsibility remains with the operator.