A hidden mark placed in AI-made content so it can later be recognised as AI-made.
An AI watermark is a signal hidden in AI-made pictures, text or audio. People can't easily notice it, but software can detect it later. Google DeepMind's SynthID, for example, can even mark text by subtly steering which words get chosen. There are also open standards such as C2PA that attach a record of how, and with what, a file was made.
Watermarks aren't foolproof. Cropping or re-saving an image, or rewording a piece of text, can weaken or remove the mark. That's why places such as the EU, China and South Korea are also bringing in rules that require AI-made content to be labelled.
Think of it like this
Hold a banknote up to the light and a hidden image appears. You don't notice it normally, but it's there when you want to check. An AI watermark is a hidden mark like that, revealed by a checking tool.
For example
If an image file carries a record saying which AI tool made it, a checking tool can show you where it came from.
Common mix-ups
Mix-upIf there's a watermark, you can always tell something was made by AI.
ActuallySome AI tools add no mark at all, and each company marks things its own way, so one company's checker usually recognises only its own mark. That's why a checker finding no mark doesn't prove a person made it.
Mix-upYou can't put a watermark in text.
ActuallyGoogle DeepMind applied its text watermark to Gemini's real answers and compared about 20 million of them, finding no clear drop in quality, as it reported in the journal Nature in 2024.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- If you labelled AI-made homework, which kind of mark would last longest?
- Activity
- Each group puts two marks in a short text: a visible one (a line at the end) and a hidden one (say, the first letters of the sentences spell a word). Another group rewrites the text in their own words or copies only part of it; then check which mark survived. Link it to lesson DAY 40 on saying when AI made something, and talk about why owning up yourself is needed as well as marks.
Records about this
Source-checked moments from AI history.
Provenance labeling such as C2PA and SynthID moves from goodwill to a regulatory compliance requirement. Dec 2, 2026 is only a transition deadline for systems already on the market, not a postponement of Article 50 as a
The first case of making the labelling of AI-generated content a legal duty
The first case of provenance marking for generated content moving beyond research into large-scale deployment