AI watermarking hides a signal inside AI-generated images, audio, video or text so software can later tell that the content came from an AI system.
In plain words
A visible label can be cropped off. A watermark is built into the content itself, invisible or inaudible to people but detectable by a matching tool. For text, one method nudges the model toward a secret, pseudo-random set of words while it writes; a detector later tests for that pattern. The EU AI Act requires providers of generative AI to mark its output in a machine-readable way.
A quick example
A chatbot adds a text watermark as it writes. Later, a teacher runs a suspicious essay through the matching detector. The words look ordinary, but the statistical pattern is there, so the detector reports that the text very likely came from that model.
Where you'll meet it
- Images, audio, video and text, for example through Google's SynthID.
- Meeting disclosure rules such as the EU AI Act.
Related
Sources
- DocsSynthID, Google DeepMind · read 13 Nov 2026
- PaperScalable watermarking for identifying large language model outputs, Dathathri et al. (Nature) · read 13 Nov 2026
- PaperA Watermark for Large Language Models, Kirchenbauer et al. (arXiv) · read 13 Nov 2026
- DocsCode of Practice on marking and labelling of AI-generated content, European Commission · read 13 Nov 2026
- DocsAI Act, European Commission · read 13 Nov 2026
Checked by WhatIsAI against the sources above · 13 Nov 2026. Spotted an error? Report it.