Developer Guillaume Meyer has released an open-source project designed to remove different types of AI watermarks and provenance markers from text and files, putting new attention on how those signals behave once AI-generated content is edited or processed. Called watermarks-remover, the project is available on GitHub and includes tools for dealing with hidden text markers, statistical watermarks and file metadata. The project comes at a timely moment for AI-content tracking, with Anthropic recently announcing machine-readable marking for supported Claude models and products.
Meyer’s project does not show that Claude’s system, or AI watermarking generally, has been defeated. Instead, it provides different methods for dealing with several types of content markers, raising a practical question for companies developing provenance systems: how well do those markers remain detectable after content is changed?
What Guillaume Meyer’s Watermarks-Remover Project Actually Does
Meyer’s watermarks-remover project is an open-source collection of tools for removing or modifying different types of AI-content markers. The project is available on GitHub, where its documentation describes methods for dealing with both text-based and file-based provenance signals. The project covers several types of markers rather than relying on one method. These include invisible Unicode characters and other hidden text signals, statistical text watermarks, and metadata attached to files. The repository also lists support for formats including PNG, JPEG, SVG, PDF, DOCX, ODT, HTML and Markdown.
🧹watermarks-remover now supports watermarks from OpenAI and Gemini in addition to Claude.https://t.co/OxSdnAjEGe
— Guillaume Meyer (@guillaumemeyer) August 11, 2026
That distinction matters because an AI watermark is not always a visible label added to a piece of content. Some markers can exist as metadata inside a file, while statistical watermarks are designed to create a detectable pattern in the way text is generated. Meyer’s project uses different approaches for those different types of signals. For some markers, the process involves removing information stored in a file or text. For statistical text watermarks, the project includes a rewriting approach intended to disrupt the detectable pattern. The repository does not claim to remove every possible AI watermark. Its documentation also distinguishes between markers that can be directly removed and statistical signals that require content to be rewritten. That means the project should not be treated as a universal tool for making AI-generated material undetectable.
The timing of the project is nevertheless notable. Anthropic recently said supported Claude models can add an invisible watermark to generated text, while supported files can receive digitally signed provenance information using the C2PA standard. Anthropic’s system is therefore useful context for understanding why projects such as Meyer’s are receiving attention, but it is not the main development here. The GitHub project covers several forms of AI provenance, while Anthropic’s system is one example of how an AI company is attempting to mark content at the point of generation or processing.
There is no basis for saying the project has defeated Claude’s watermarking system. Instead, it shows how developers can experiment with removing or altering different types of markers once content is outside the platform that produced it. The project also raises a broader question about what happens when marked content is edited, translated, rewritten or converted between formats. If a provenance signal changes during that process, the usefulness of the original marker can depend on how reliably it can still be detected.
How Anthropic’s Claude Watermark Works and What It Proves
Anthropic says Claude models launched in the European Union on or after August 2, 2026, support machine-readable marking from launch, with the system also applying to supported models and products elsewhere. For generated text, Claude embeds an invisible watermark into its output. For supported files, the company can add digitally signed provenance metadata using the C2PA standard. Anthropic’s explanation of how Claude marks AI-generated content explains that the text watermark is different from a conventional AI detector. Instead of examining a finished article and trying to estimate whether it was written by a person or an AI model, the mark is added when Claude generates or processes the material. A detection system can then look for that embedded signal.
That distinction matters because a watermark does not necessarily establish authorship. Someone could write an article themselves and ask Claude to edit, summarize or translate it. If Claude processes that material, the resulting text could carry a Claude mark even though the person supplied the original work. Anthropic also says the absence of a detectable watermark should not be treated as proof that AI was not involved. Heavy editing, paraphrasing and translation can make the signal harder to detect, while short passages can provide less information for reliable detection. The company describes the marking system as a way to provide information about the origin of content rather than a definitive test of whether a human or AI was responsible for every part of it. That distinction is particularly relevant to Meyer’s project because it shows why removing or disrupting a marker is not necessarily the same as proving that AI was never involved.
Meyer’s watermarks-remover covers several types of AI-content markers, and different forms of provenance information require different methods to remove or alter them. Different forms of provenance information require different methods to remove or alter them, and changes to content can affect whether a watermark remains detectable. The two developments therefore sit on opposite sides of the same technical question. Anthropic is adding a machine-readable signal to help identify content that Claude has generated or processed, while Meyer’s open-source project gives users tools for examining and modifying different types of AI markers.
You may also like to read – OpenAI Launches GPT-5.6-Cyber for Vetted Security Researchers
