Glossary
What is an AI watermark?
In short
An AI watermark is a hidden mark that a neural network adds to the text, image, sound, or video it creates, so software can later detect where the content came from.
Security and law · AI watermark
MMOOZZGG.com editorsChecked: 1 min read
For a child
Real banknotes have a picture that you see only when you hold the note up to the light. It shows that the note is not fake. Some smart programs that draw pictures and write texts also hide an invisible sign in their work. You cannot see it with your eyes, but a special check finds it and says: a computer made this.
In plain words
When a neural network creates a picture, a sound, or a video, it can change the pixels or the sound wave slightly, so a person sees no difference but a checking program finds it. In text the mark is hidden differently: the model chooses words from a set defined in advance a little more often, and their share shows that the text was generated. The mark helps tell what a neural network created from what a camera filmed or a person wrote. It has a weakness too: heavy editing, retelling, or filming the screen can erase it, and usually only the party that set the mark can check it.
For business
Marks in generated content are becoming a duty. The EU AI Act requires providers of generative systems to mark output in a machine-readable form, and its transparency rules apply from August 2026. Russia’s law 243-FZ speaks of a notice that AI was used, not of a hidden mark. For a company, this raises two questions. If you publish pictures and videos from a neural network, check whether the service adds a mark and whether your editing removes it. If you check other people’s content, remember that a missing mark does not prove a person made it, because many generators add none. Write the process for such content into your rules for using generative AI.
For a specialist
A watermark for generative AI embeds a signal in the output that people cannot notice and an algorithm can detect. For text, Kirchenbauer et al. (2023) proposed a scheme with a “green” list: before each token is generated, the vocabulary is split pseudo-randomly into two parts, the model slightly raises the probability of the green part, and a detector tests a statistical hypothesis by the share of such tokens without access to the model’s parameters. For images, sound, and video, the signal goes into the pixels or the spectrum; this is how SynthID works, which, as Google DeepMind describes it, marks images, audio, text, and video. The properties of a scheme are invisibility, resistance to compression and edits, a low false positive rate, and payload size. A watermark is different from content provenance metadata: the C2PA 2.1 specification (September 2024) links a signed manifest to the file through cryptographic hashes and allows a watermark as a soft binding. Paraphrasing and re-encoding weaken the mark, so it is seen as one signal, not as proof.
Example
A marketer got photos for an ad from a contractor and is not sure that a photographer took them. A check finds a generator’s mark, and the company agrees with the contractor on how to caption the images in the publication.
Related terms
Questions and answers
Sources
- Kirchenbauer et al., A Watermark for Large Language Models, 2023arxiv.org
- Google DeepMind, SynthID, checked 16 October 2026deepmind.google
- C2PA, Technical Specification 2.1, September 2024spec.c2pa.org
- Regulation (EU) 2024/1689, Article 50, AI Act Service Desk of the European Commissionai-act-service-desk.ec.europa.eu
- European Commission, AI Act: regulatory framework on AI, application timelinedigital-strategy.ec.europa.eu
Not sure whether your company needs this?
AI solutions map