AI watermark

in English: AI watermarking

basicspreading

An AI watermark is a signal that is invisible to the eye but machine-readable, which an artificial intelligence system embeds in the images, audio, video or text it creates.

In plain terms

If you hold a banknote up to the light, a faint image appears that verifies its authenticity. An AI watermark plays a similar role in digital content: a signal is embedded in machine-generated images, audio, video or text that you do not notice, but a verification program does.

It is useful for determining after the fact that content was created by artificial intelligence. This helps filter out fake news, false videos, and misleading ads, and enables platforms to label content as "AI-generated".

Unlike a logo or caption placed in the corner of a photo, which anyone can see and easily crop out, a hidden watermark is scattered within the content itself, in the image pixels or the text's word choices, so it usually survives simple resizing or compression.

It is not a magic weapon. It only marks content created by a watermarking system, which can be weakened or removed with clever manipulation.

How it works

Embedding: the generative model hides a tiny, regular pattern in the content while creating the image, sound, or text, or immediately afterward.
In images and audio, this refers to tiny modifications of pixels or sound waves that the human eye and ear do not detect.
With text, the model selects certain words slightly more frequently based on a secret key. This is not noticeable in a single sentence but is statistically detectable in longer texts.
Recognition: the detector program, knowing the key, finds the pattern and states the probability that the content contains a watermark.
Watermarks are often used together with signed provenance data (metadata) because the two complement each other: metadata is detailed but easily stripped, while watermarks are concise but more persistent.

An everyday example

You scroll through a social media site, and under a startling photo of a politician a "Made with AI" label appears. Among other ways, the platform can know this because the service that generated the image embedded an invisible watermark or signed provenance data, and a program read it upon upload.

Who is affected

everyoneIn news and social media, images or videos are increasingly labeled as machine-generated based on watermarks or attached provenance data.

marketerAI-generated campaign materials from major providers often carry watermarks or provenance data, and platforms can automatically flag them based on this. → Marketing

parent, student, teacherTextual watermarks could in principle help filter out machine-written essays, but no one can be fairly penalized based solely on this. → Education & Family

developerIf you offer an agent in the EU that produces generative features or content, the AI Act requires, as a general rule, that from August 2026 the output be marked in a machine-readable way, so you need to plan for it from the design stage. → Code

A common misconception

MythIf content has no AI watermark, it was definitely made by a human.

In realityThe absence of a watermark proves nothing: many models do not watermark at all, and the signal can disappear through transformation. The presence of a watermark is a strong indicator, but its absence is not proof.

Limits and risks

A watermark can be weakened or even removed by aggressive cropping, re-encoding, screenshots, or rewriting the text.
There is no unified watermark standard: one company's detector typically does not recognize another company's watermark.
With open, locally runnable models, watermarking can be disabled, so malicious users are not stopped.
Recognition is probability-based, so it can err in both directions, especially with short texts.

Similar, but not the same

AI-detektor – The detector tries to guess afterward, based on style, whether a machine wrote the text. A watermark, by contrast, is a deliberately placed signal created at the time of generation, so if present, it is more reliably detectable.
Content Credentials (C2PA) – This attached, signed 'provenance sheet' documents the history of the content, while the watermark is embedded within the content itself and may persist even if accompanying data is lost.
Visible watermark – A logo or caption is meant for humans and can be easily cropped, while a hidden AI watermark is for machines and invisible to the naked eye.

Related concepts

Broader

Digital watermark, Content provenance verification, Generative AI

Narrower

SynthID, Text watermark, Audio watermark

On the edge

Deepfake, AI-detektor, C2PA, Synthetic media, EU MI-rendelet

A brief history

Digital watermarking emerged in the 1990s to protect copyright. The most-cited method for watermarking text from large language models appeared in 2023. Google DeepMind presented its SynthID system the same year, first for images and then, in 2024, for text. The EU's AI Act, adopted in 2024, requires machine-readable marking of machine-generated content from August 2026.

Where things stand

Major providers' image, audio, and video generators typically already embed watermarks or signed provenance data in their output, and since August 2026, EU law has required machine-readable labeling. There is not yet a unified watermark standard that works across companies. Text watermarks are less widespread and can easily be weakened by paraphrasing.

Status: 2026 November

In the news

Appears in 5 Agent Age stories; the curve shows weekly occurrence in the news magazine.

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
Anthropic to add invisible watermarks to text generated by ClaudeAug 17, 2026 · Agent Age
Anthropic's Claude to add invisible watermarks and metadata to AI-generated content under EU rulesAug 11, 2026 · Agent Age
European transparency rules on labelling AI text and content affect marketersAug 10, 2026 · Marketing

Also known as

AI watermark · AI watermark · AI watermarking · Invisible watermark · AI watermark · AI watermarking

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.