The EU AI Act Art. 50 labeling rule, explained
This page explains the rule in plain language. It is not legal advice - if you need a definitive answer for your situation, talk to a lawyer.
What Article 50 requires
Article 50 of the EU AI Act requires that AI-generated or AI-manipulated image, audio, and video content ("deepfakes" and similar synthetic media) be marked in a way that makes clear it was produced or altered by AI. The marking has to be machine-readable and detectable, and - where technically feasible - effective, interoperable, robust, and reliable. Article 50 has applied since 2 August 2026.
The article actually sets two separate duties: Article 50(2) is a machine-readable marking duty on providers - whoever builds and operates the generative AI system. Article 50(4) is a separate, visible disclosure duty on deployers - typically professional or business users who publish deepfake-like content to the public. This tool is neither: it's a self-service aid that lets you add the marking yourself to an image you already have, regardless of which system generated it.
An artistic, satirical, or fictional work still needs disclosure - the exemption only lightens how obtrusive that disclosure has to be (for example, a subtle watermark instead of a prominent banner), it doesn't remove the obligation.
Who it applies to
The obligation falls on providers of AI systems that generate synthetic image, audio or video content, and on deployers of such systems in certain cases (for example, deployers who publish a deepfake). It does not generally require every individual sharing an AI image online to add a label themselves - but where you are acting as the deployer of a generative AI system, or you want to be transparent regardless, marking the content is the safer choice. Non-compliance penalties (up to EUR 15 million or 3% of global annual turnover, whichever is higher) attach to the providers and deployers who carry the Article 50 obligation - not to someone using this tool to label their own image.
What counts as "AI-generated" vs. "AI-modified"
The common distinction used by metadata standards (and by this tool) is:
- AI-generated - the image was produced entirely by an AI model, with no real photograph as a starting point.
- AI-modified - a real photograph was altered using an AI tool, for example inpainting, outpainting, or generative fill.
How this tool helps
This tool does two things to an image you upload: it draws the official EU AI icon onto it, and - for PNG and JPEG - it embeds an IPTC Digital Source Type value in the file's XMP metadata, recording whether the image is AI-generated or AI-modified. Both the icon and the metadata are a form of the "machine-readable marking" Article 50 talks about.
What it does not do: this isn't C2PA / Content Credentials. The XMP metadata it writes is plain and unsigned, not a cryptographic signature, so it's weaker evidence than a C2PA credential and can be stripped just by re-saving the file, re-uploading it to most platforms, or taking a screenshot. Using this tool also doesn't by itself make your content "compliant", and using the EU AI icon isn't a Commission certification or endorsement - whether a given piece of content needs labeling, and whether this labeling is sufficient for your situation, is your call to make. See our guide on verifying the embedded metadata to confirm the label actually made it into your file, or use our Check page to read a file's existing metadata back out. Check only looks for our kind of label (and whether a C2PA marker is present) - it doesn't analyze pixels, so a file with no label found was not necessarily made without AI.