Watermark

If you hold a banknote up to the light, you can see a faint pattern in the paper. This is exactly where the word watermark comes from. In the digital world, it means something similar: a marker is embedded in a file that normally goes unnoticed. This marker indicates who created the file, or with what. It can later be read out again using a verification program. There is an important difference from a visible logo in the corner of an image: anyone can crop out a logo, but a hidden watermark is spread across the entire file.

Why forgeries hardly stand out without a marker

Programs that generate images, voices, and texts on command have become very good. A fabricated photo of a politician can be ready in a minute. Often it can no longer be distinguished from a real photograph by eye. Watermarks are meant to provide a kind of proof of origin here.

This affects several areas at once. In elections, it concerns fake videos and imitated voices on the phone. In schools and universities, it concerns texts that were not written by the student themselves. Photographers and musicians, in turn, want to prove that a work originates from them and was not copied by a machine.

Politics has also picked up on this point. The European Union’s AI regulation, often called the AI Act, requires providers to label artificially generated content in a machine-readable way. Watermarks are one of the most important technical means for this. For large providers, this is therefore no longer a voluntary gesture, but an obligation.

How a marker hides in the noise

An image consists of millions of individual pixels, each with a color value. A watermark alters many of these values by tiny amounts. A shade of red becomes one step darker, its neighbor one step lighter. To the eye, the image remains the same. The verification program, however, knows the embedded pattern and finds it again.

With text, this works differently, since there are no color values to adjust. A text program selects each next word from several almost equally good options. The watermark slightly influences this selection: a certain half of the permitted words is given a slight preference. A single sentence proves nothing on its own. But across a whole page, a statistical pattern emerges that can hardly be coincidental.

The weak point of all these methods is durability. An image does not survive cropping, shrinking, and heavy re-saving undamaged, and the watermark can get lost in the process. With text, it is sometimes enough to rephrase the content. A second misconception is widespread: if the watermark is missing, the file is not automatically genuine. Anyone wanting to deceive simply uses a program without labeling.

Watermarks in images, videos, and streaming services

Google marks images from its image programs with the SynthID method. It works invisibly and can be queried with a verification tool. Other providers additionally rely on so-called metadata, i.e. additional information in the file header that records camera, date, and editing steps. The industry standard for this is called C2PA. Both complement each other: metadata is easier to read, watermarks are harder to remove.

Social networks such as Instagram, TikTok, and YouTube now check uploaded files for such markers. If one is found, a note such as “AI-generated” appears beneath the post. So you don’t see these notes because a human checked the video, but because a machine found a pattern.

However, watermarks have existed long before the debate about artificial intelligence. Streaming services embed an invisible customer ID into every film. If a pirated copy surfaces, the leak can be traced back. The same basic idea is found in banknotes, ID cards, and concert tickets. In business news, you usually encounter the term in connection with new requirements for AI providers.