How AI content labels and watermarks work, and what the EU now requires

Three tools are used to show that an image, a video, an audio clip or a piece of text was made with AI. An invisible watermark hides a signal inside the content itself. Signed provenance metadata, known as Content Credentials, records where a file came from and how it was changed. A visible label tells the viewer directly. Each catches different things, none is foolproof, and since August 2, 2026, EU law has required them in specific cases.

Watermarks: a signal hidden in the content

Google's SynthID is one example. Google says it embeds digital watermarks directly into AI-generated images, audio, text or video, and that the marks are imperceptible to people but detectable by SynthID's own technology. For images and video, the watermark is added the moment content is created and is designed to withstand cropping, filters, changes in frame rate and lossy compression, according to Google.

Text works differently. A language model writes one word, or token, at a time, and each candidate word has a probability score. SynthID nudges those scores so that the pattern of word choices carries a watermark, which Google says does not affect the quality of the output.

Google is open about the limits. Its developer documentation says watermarking “is less effective on factual responses,” because there is less room to vary the wording without hurting accuracy. It also says detector confidence “can be greatly reduced” when AI-generated text is thoroughly rewritten or translated into another language. Google has released SynthID Text as open source so that other developers can use it.

Detection depends on who made the mark. Google's SynthID Detector checks whether an image, video or audio file was made with AI from Google or from partners it names, including OpenAI, NVIDIA and Kakao.

Content Credentials: a signed record of origin

The Coalition for Content Provenance and Authenticity (C2PA) publishes an open technical standard called Content Credentials. A credential, also called a C2PA manifest, holds statements about a file: when and where it was created, what tools modified it and how AI was used. The manifest is signed with the private key of the software or hardware that did the work. Because the signature is tied to a cryptographic hash of the file, any change to the file or to the record breaks the match, which makes tampering evident.

Two limits matter. First, the record can be removed. Asked whether provenance metadata can be stripped from a file, the C2PA's own explainer answers: “Yes it can.” Its fix is a “durable” credential that pairs the signed record with a watermark or a fingerprint, so a removed credential can be found again in cloud storage. Second, a valid credential shows that the record is intact and who signed it. It does not show that the picture is true. The C2PA says provenance information “alone cannot tell you whether the digital content is true, accurate or factual.”

What EU law requires since August 2, 2026

Article 50 of the EU AI Act sets transparency rules that apply from August 2, 2026. Providers of AI systems must apply a machine-readable mark to synthetic content their systems generate or manipulate, and make it detectable. Tools that only assist with standard editing, or that do not substantially change the input, are exempt.

Deployers of AI systems must clearly label deepfakes: AI-made images, audio or video that resemble real people, objects, places or events and would falsely appear authentic. They must also label AI-generated text published to inform the public on matters of public interest, unless a person has reviewed it and someone holds editorial responsibility for it.

Generative AI systems placed on the market before August 2, 2026, have a transition period for the marking rules until December 2, 2026. Deepfakes made before August 2 do not have to be labeled retroactively, though the Commission encourages it. Companies that break the transparency rules face fines of up to €15 million, or up to 3% of total worldwide annual turnover.

To help companies comply, independent experts drew up a voluntary Code of Practice on Transparency of AI-generated Content, published in June 2026. The Commission and the EU's AI Board have confirmed that it is an adequate way to demonstrate compliance, and about 190 organizations had signed it by the end of July 2026. The EU also offers three optional icons that deployers can use as labels. Companies that do not sign must show by other means that their marking and labeling is adequate.

What this means for readers

A missing watermark or credential does not prove that something was made by a person. Watermarks can weaken when content is edited, and metadata can be stripped. The C2PA itself warns that no one should judge a file's trustworthiness purely on whether it carries Content Credentials. That is why its durable credentials combine signed records with watermarks or fingerprints, rather than relying on one method.

The practical test is to check several signals: a visible label, a credential that names a signer you recognize, and a watermark check where a detector exists. The next AI Act deadlines, for high-risk systems, come in 2027 and 2028; we covered the first harmonized standard for those rules.