For images and audio, a watermark can be a pattern spread across pixels or sound that people cannot perceive but a detector can. Text is harder because there are no pixels to hide a pattern in. One well-known approach, from Kirchenbauer and colleagues, nudges the model toward a secret, pseudo-random subset of words at each step; a detector that knows the secret can then spot the statistical tilt in a short passage, while a reader sees ordinary text. Google DeepMind’s SynthID applies watermarking across images, audio, text and video, and its text method was published in Nature in 2024.
A concrete use: a news platform that runs a detector on uploaded images can label pictures made with a watermarking generator before they spread. The catch is that this only works for content from tools that embed a watermark and share a detector, and heavy editing or paraphrasing can weaken the signal. A missing watermark does not prove content is human-made.
Regulation is pushing adoption. Article 50 of the EU AI Act requires providers of generative systems to mark outputs in a machine-readable, detectable way where technically feasible. Editorially, watermarks are best paired with content provenance: the watermark is embedded in the content itself, while provenance metadata records who made it and how it was edited but can be stripped. Both serve AI transparency.
Sources
- Kirchenbauer et al.: A Watermark for Large Language Models — Proposes embedding signals in generated text that are invisible to humans but algorithmically detectable from a short span of tokens, with negligible impact on text quality; checked October 24, 2026.
- Dathathri et al.: Scalable watermarking for identifying large language model outputs (Nature, 2024) — Describes SynthID-Text, Google DeepMind’s production text-watermarking scheme; checked October 24, 2026.
- Google DeepMind: SynthID — Says SynthID embeds imperceptible digital watermarks directly into AI-generated images, audio, text or video that can be detected by software; checked October 24, 2026.
- EU AI Act Explorer: Article 50, Transparency Obligations for Providers and Deployers of Certain AI Systems — Requires providers of generative systems to mark outputs in a machine-readable format detectable as artificially generated, as far as technically feasible; applies from 2 August 2026; checked October 24, 2026.