AI companies are increasingly looking for ways to mark content generated by their models. Now, someone has built an open-source tool designed to remove some of those marks.
A GitHub project called watermarks-remover is designed to strip different types of AI provenance signals from text and files. According to its documentation, it can work with invisible Unicode characters, statistical text watermarks and metadata embedded in formats including PNG, JPEG, SVG, PDF, DOCX, ODT, HTML and Markdown.
That sounds straightforward until you look at how some of these watermarks actually work. A metadata tag can be removed relatively cleanly. A statistical watermark embedded into the way an AI model chooses words is a different beast altogether. And that is where this project gets interesting. And before we move ahead, the developer has stated the new tool now works with OpenAI and Gemini, alongside Claude.
Some AI watermarks are buried in the way one writes
The project separates its approach into different layers. One targets edit-based signals such as unusual Unicode characters, while another focuses on statistical patterns in generated text. A separate cleanup layer deals with file provenance information such as C2PA, EXIF, XMP, and document properties.
The latest version 0.3.1 makes the text rewriting side more aggressive. Instead of simply swapping a few words, the tool can change sentence structure, word choices, transitions, and other patterns in an attempt to disrupt statistical watermarking. It also includes options intended to make rewritten text sound more natural.
There is an obvious catch, though. Rewriting text to remove a statistical watermark can also change the text itself. The project’s documentation admits that this process can affect tone, voice, and precision, particularly when a large portion of the original wording needs to be changed.
So this isn’t quite the digital equivalent of hitting “remove watermark” in an image editor. In a post on Wednesday, Meyer said that his tool only gets rid of metadata right now. He said that getting rid of the real marks might come later, but not today. His README goes even further, saying that a rewrite changes the words used by the premium model to those used by the cheaper model. It also asks why someone would pay more for a premium model and then use a worse model to process its output.
The developers aren’t claiming this will fool every AI detector
In fact, the project is pretty clear about its limitations. It describes the rewriting process as best effort and says it cannot guarantee that a particular vendor’s detection system will fail. Some signals can also remain after the cleanup process.
That matters because there isn’t one universal AI watermark hiding inside every piece of generated content. Different companies and systems can use different approaches, and the project itself divides those signals into several categories. The repository also says its purpose is privacy and research, rather than helping people falsely claim that AI-generated work was written entirely by a human.
For now, watermarks-remover is best understood as a sign of where the AI industry is heading. Companies are looking for ways to establish provenance and identify AI-generated material, while developers are already exploring how to remove or disrupt those signals.
The interesting part isn’t whether this particular GitHub project can beat every AI detector. It probably can’t. The interesting part is that AI watermarking is already becoming a cat-and-mouse game, and we’re still very early in it.
