Claude's Invisible Watermarks: How Claude Marks Your Text and Images
Since August 2026, Claude signs what it generates with invisible marks, one in text, one in images. Here is how it works, explained simply, plus the skill to clean content you own.
August 29, 2026 · 6 min read
Since August 2, 2026, Claude no longer just generates text and images: it signs them. Every piece of content produced by a new model carries an invisible mark, one you never see while reading, but one a machine can find. It is a hidden signature, not a visible stamp, and it changes nothing about what you read.
There are two different marks, and they do not work the same way: one for text, one for images and files. Most people confuse the two, and above all confuse the real mark with a myth that circulates everywhere. This guide explains both as if you knew nothing about it, then hands you the tool to clean content you own.
The Claude AI Lab is my Skool community where I share my Claude systems and the more advanced modules. Access is $67/month.
Join the Lab →The mark in the text
Picture Claude writing a sentence. At each word, it often has several options that are equally good: "quick", "fast", "rapid". To choose, it secretly follows a pattern, a kind of rigged coin flip whose rule only the company knows. Word after word, these tiny choices draw an invisible rhythm.
Later, to check whether a text came from Claude, you just re-read those word choices and see if they follow the secret pattern. The mark is not an ink or a hidden letter: it is a habit in the way words are chosen.
Claude has not published the exact algorithm, but this is the known family of these watermarks: the choice of words is nudged according to a secret pattern. The mark survives a copy-paste and a light edit, not a full rewrite.
The mark in the images
For images and files, Claude switches methods entirely. There, it does not touch the content: it sticks a signed label onto the file, reading "this file was produced by Claude, on this date." That label follows an industry-shared standard called C2PA.
The label is sealed: if someone changes it, it shows. So it is reliable proof as long as it is there. The catch is that a glued label peels off.
The text mark is IN the content (the choice of words), so it travels with it. The image label is GLUED to the file, so it comes off the moment the file is rebuilt. That is why an image watermark is far easier to lose than a text watermark.
The invisible-character myth
Here is the most common mistake. People paste AI text into a tool that displays hidden characters, land on strange invisible spaces (thin spaces, zero-width spaces), and call victory: "I found the watermark!". No.
Those ghost characters do exist, but they are formatting debris, not the signature. The real text watermark lives in the choice of words, and the choice of words shows up in no character inspector. Anthropic says it plainly: the mark is woven into the text itself, not added as metadata or hidden characters.
Deleting those invisible characters gives you the feeling of having removed the watermark, while you touched nothing of the real mark. Two different things live in the same place, and you need to know which one you are aiming at before you clean.
The skill that removes the marks
Now that you know where the marks live, here is the tool that inspects and removes them, on content you own. It is an open-source skill for Claude Code, remove-ai-marks, in Guillaume Meyer's watermarks-remover repository.
You install it from Claude Code in two commands:
/plugin marketplace add guillaumemeyer/watermarks-remover
/plugin install watermarks-remover@watermarks-remover
The repository and its full documentation live at github.com/guillaumemeyer/watermarks-remover.
The tool is meant for your own content: privacy, hygiene, research. Cleaning what you own is legitimate. Using it to pass a piece of work off as something it is not is not, and no tool makes that true.
Understanding these marks is mostly about knowing what travels with your content when you publish. The text mark follows the words, the image label follows the file, and the invisible characters are just a false lead. The rest is hygiene: you inspect, and you decide.
Install the skill, run it on a text or image you generated yourself, and read its inspection report first. You will see concretely what is marked, before you clean anything at all.
Want to go further?
In the Lab, I share the Claude resources I actually use and how I wire them into my systems.
A dedicated session or program, tailored to your tools and use cases.
And day-to-day, I post one reel a day on Instagram: @quentin_iamarketing