The EU AI Act and labeling of AI-generated content
What the EU AI Act is
The EU AI Act (the European Union’s Artificial Intelligence Regulation) is the legal framework with which the EU regulates the use of AI. Among its many goals, one affects anyone who publishes images directly: pushing transparency about content generated or manipulated with artificial intelligence.
The idea is simple to state: if content is synthetic, the public should be able to know. That is where the pressure toward labeling AI-generated content comes from.
What it promotes on labeling
In broad terms, the regulation pushes generative AI systems and the people who deploy them to make synthetic content identifiable. That includes the expectation that AI-generated images, audio and video be marked so they can be recognized, for example through technical signals embedded in the file itself.
It is not an isolated phenomenon: it lines up with the industry’s move toward provenance standards. The most widespread is C2PA / Content Credentials, which we explain in depth in C2PA and Content Credentials. In practice, these credentials are one of the technical mechanisms the ecosystem uses to answer this demand for transparency.
This article is a plain-language explanation of the regulatory context, not legal advice. Concrete duties depend on your role (provider, deployer, user), the use case and the regulation’s application timeline.
Why this reaches your photos
The practical effect is that large platforms have adopted the reading of provenance metadata to label content automatically. When you upload an image, the platform checks whether it carries C2PA credentials, XMP fields or other marks that indicate AI involvement, and applies the AI label accordingly.
This has an uncomfortable and honest consequence: the system does not distinguish well between:
- A fully synthetic image, created from scratch with AI.
- A real photo that received a small retouch with a tool that writes provenance credentials (Photoshop, Lightroom, Canva).
The result is false positives: authentic photographs that receive the AI notice because their file contains provenance marks, not because the scene is fake.
Transparency, yes — but with control over your data
The goal of labeling — giving the public transparency — is reasonable. That does not mean you have to publish everything your files hide. The same metadata that triggers the label often also includes private information: GPS location, camera model, exact date. We talk about that in metadata and privacy.
You get to decide what you share. The recommended flow is:
1. Check what the file carries with the checker: C2PA/JUMBF credentials, EXIF, IPTC, XMP and PNG text chunks. 2. Clean what you do not want to publish, with specific modes such as /c2pa, /exif and /xmp.
Cleaning is done without recompressing or changing the pixels (verifiable by hash), with no added watermark, no account and for free. The file is processed on the server and deleted after 5 minutes. We never claim the process happens “only in your browser”: the image does go to the server.
The limits of cleaning metadata
It is important to be transparent about how far this goes, especially in a regulatory context that pushes labeling from several directions:
- Removing provenance metadata does not remove invisible pixel-level watermarks, such as Google’s SynthID. Those live in the image, not in the metadata.
- Platforms can detect synthetic content by means other than metadata, including automatic classifiers.
If an image was truly generated by AI and carries a pixel-level watermark, cleaning metadata will not make it “undetectable.” We prefer to say that clearly.
In short
The EU AI Act pushes labeling of synthetic content, and standards such as C2PA are the industry’s technical response. That same mechanism causes real photos to receive the label because of provenance metadata. You can check and clean your files to control what you publish — both for the label and for your privacy — knowing that pixel-level watermarks are outside that scope. Expand the concepts in the glossary.