The EU's AI Labeling Rules, Explained for Normal People

Since August 2026, the European Union requires AI-generated content to be labeled. If you have seen "AI-generated" tags popping up on images and videos, this is why. Here is what the rules say and what they change for you.

What happened

Article 50 of the EU AI Act took effect on August 2, 2026. It is the transparency part of the law. The core idea is simple: when content is AI-generated, people should be able to tell.

The rules put duties on two groups. Providers, the companies that build AI systems, must make sure their outputs carry machine-readable marks so they can be detected as AI-generated. Deployers, the businesses and organizations that use those systems, must clearly label deepfakes and certain AI-generated text published on matters of public interest.

Violations can draw fines of up to 15 million euros or 3 percent of a company's global annual revenue, whichever is higher.

What "machine-readable" means in practice

The law does not ask for a visible stamp on every image. It asks for marks a machine can read. In practice that means two technologies you may have heard of: C2PA Content Credentials, a signed metadata record, and invisible watermarks like Google's SynthID that live in the pixels.

The EU published a final Code of Practice in June 2026 that points at C2PA as the machine-verifiable layer. One notable admission in that document: no single marking technique currently meets the law's bar for effectiveness and reliability on its own. The Code therefore expects a layered approach, using at least two marking methods where necessary. That is why you see the big labs doing both at once. OpenAI, for example, now embeds Content Credentials and SynthID watermarks in ChatGPT images together.

The law also expects providers to offer detection mechanisms, so that users, journalists, and researchers can verify whether content came from their AI systems.

The deadlines are not all the same date

This trips people up. August 2, 2026 was the headline date, but the obligations phase in:

• Chatbot disclosure and the general labeling duties took effect August 2, 2026.
• The technical watermarking duty was pushed to December 2, 2026 under a legislative package called the AI Omnibus.
• AI systems already on the market before August 2026 get a transition period, with full application by December 2026.

So if you are wondering why labeling still looks patchy months after the "deadline," that is why. The full regime is still phasing in.

The exceptions

Not everything AI-touched needs a label. The law carves out cases that matter for everyday life:

• If AI-generated text went through genuine human review and someone takes editorial responsibility for it, the labeling duty does not apply. A newspaper that uses AI drafting but edits and stands behind the piece is fine.
• Artistic, creative, satirical, and fictional works get lighter treatment. A meme or a short film is not held to the same standard as a fake news photo.
• Chatbots do not need to announce themselves when it is obvious from context that you are talking to a machine.

What changes for you

Honestly, less than the headlines suggest, and in a good way. You do not have a legal duty to label anything. The duties fall on companies. What you get is more information in the wild: more "AI-generated" labels on content, an official set of EU icons for marking synthetic media, and provenance info showing up in places you already look. Google has said C2PA verification is coming to Search and Chrome.

The most misunderstood part: a missing label tells you nothing. The rules are still phasing in, most AI tools mark nothing at all, and labels get stripped whenever files are re-uploaded or screenshotted. The law makes the labeled world more transparent. It says nothing about the unlabeled world.

That is the same reason our image checker answers "cannot tell" when a file carries no signals. Absence of a label is not evidence of a human behind the camera. The EU's own framework is built around that exact asymmetry: a mark you can verify means something, and the lack of one means you keep checking.