watermarking

basicspreading

Watermarking is the placement of a hidden, machine-readable signal in text, image, audio, or video created by artificial intelligence, so that its origin can be identified later.

In plain terms

If you hold a banknote up to the light, a faint image appears in the paper: this is the watermark, which proves the money's authenticity. AI watermarking transfers this same concept to the digital world. The difference is that here the signal indicates whether the content was generated by a machine or which system it originated from.

A good watermark is invisible to the naked eye, inaudible to the ear, and does not degrade content quality. Only a dedicated checking program can detect it. This differs from a caption or logo placed in the corner of an image, which anyone can see and cut off with a single move.

It is useful for making it easier to filter out misleading content, such as fake photos, fake news, or cloned voices. Platforms, editorial teams, and authorities can thus decide more quickly what they are facing.

It is important that the watermark does not say whether the content is true; it only indicates its origin.

How it works

In the case of images, audio, and video, the system makes tiny modifications to pixels or sound waves that are imperceptible to humans, and together these form a hidden pattern.
With text, the language model subtly shifts word choice: based on a secret key, it selects certain words slightly more often than others. This is not visible in a single sentence, but is statistically detectable in longer texts.
The signal travels with the content: it usually remains even if you download the image, forward it, or copy the text unchanged elsewhere. However, strong transformation or rephrasing can weaken it.
The checking program that knows the key examines the content and estimates how likely it is that the pattern is present. It gives no definitive verdict, only a probability.
Watermarks are often used together with signed metadata and visible labels, as neither method is reliable enough on its own.

An everyday example

On a social media site, a shocking photo appears with the caption: "Made with AI". The platform does not always know this from the uploader's admission. Sometimes the image generator places a machine-readable signal in the image, such as a hidden watermark or origin data attached to the file, and the platform's system recognizes this.

Who is affected

everyoneYou are encountering more machine-generated content in news and social media, and watermarks are one tool that helps distinguish them from real recordings.

parent, student, teacherIn schools, it often comes up whether an essay was written by a machine, but text watermarks still do not provide conclusive proof today. → Education & Family

marketerIf you generate images, video, or audio with AI, expect the content to be marked, and platforms may apply labels based on this. → Marketing

lawyer, citizenArticle 50 of the EU Artificial Intelligence Act, applicable from August 2026, requires providers of generative systems to label machine-generated content in a machine-readable manner. It does not prescribe a specific technique; watermarking is just one possible solution. → Law

A common misconception

MythIf content lacks a watermark, it was certainly created by a human.

In realityThe absence of a watermark proves nothing: many systems do not label at all, and the mark can be removed or destroyed. Only a found watermark reveals something; a missing one does not.

Limits and risks

The watermark is vulnerable: strong compression, resizing, or photographing of the image, as well as rewriting or translating the text, can weaken it or even remove it.
There is no unified system: each developer can reliably read only its own watermark, and the marking can be switched off in open-source models.
Verification can err in both directions, so do not base a serious decision, such as an accusation of fraud, solely on this.
The method barely works with short text because a few sentences do not contain enough words to reveal a statistical pattern.

Similar, but not the same

content provenance (C2PA) – There, a signed metadata record attached to the file describes the content's origin and edits, while the watermark is woven into the content itself.
MI-detektor – The detector tries to guess after the fact, without a pre-placed signal, whether a machine wrote the text, so it makes mistakes far more often.
visible label – The caption or logo is aimed at humans and can easily be cropped, while the watermark is hidden and designed for machine verification.

Related concepts

Broader

transparency, content provenance verification, responsible AI

Narrower

text watermark, image watermark, audio watermarking

On the edge

deepfake, generative AI, MI-detektor, disinformation, EU MI-rendelet

A brief history

Watermarks pressed into paper began to be used in Italian paper mills at the end of the 13th century, while digital watermarks emerged in the 1990s for copyright protection of images and music. A highly influential scientific proposal for watermarking text from language models was made in 2023. Google DeepMind presented its SynthID solution in the same year, which first marked images and later text, audio, and video.

Where things stand

By the end of 2026, the image, audio and video generators of several major providers already embed hidden watermarks or signed provenance data in their output. The EU AI Act's labeling obligation has applied since August 2026, with a transition period for systems placed on the market earlier. However, there is no common watermark standard that all vendors can read, and text watermarks remain easy to circumvent.

Status: 2026 November

In the news

Appears in 12 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Anthropic watermarks four older Claude models under EU AI Act; xAI to enable Grok API text watermarkingOct 31, 2026 · Agent Age
Parents use AI to remove watermarks from school photos to save £25Oct 6, 2026 · Education & Family
ChatGPT text to carry invisible watermark in the European Union, Hungarian outlet reportsOct 6, 2026 · Education & Family
OpenAI adds invisible watermarks to ChatGPT and Codex text in the EUOct 5, 2026 · Code
OpenAI introduces text watermarking in the EU to comply with AI ActOct 5, 2026 · Jog
AI detectors prove unreliable after paraphrasing; watermarking raises concerns for non-native writersSep 7, 2026 · Education & Family
Anthropic to add invisible watermarks to text generated by ClaudeAug 17, 2026 · Agent Age
Claude to watermark generated code and text, and developers push backAug 13, 2026 · Code

Also known as

watermarking · watermark · digital watermark · AI watermark · AI watermark · watermarking · AI watermark · SynthID

Sources

Kirchenbauer et al.: A Watermark for Large Language Models (arXiv, 2023)

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.