Watermarks Remover: Guillaume Meyer's Open-Source Tool Goes Viral

Watermarks Remover, an open-source project by Guillaume Meyer, aims to disrupt invisible AI watermarks through iterations on text and images. Following a post on X on August 11, he claims over 2 million impressions and viral engagement across multiple networks. Meyer supports content attribution but criticizes watermarks based on statistical signals, which he believes are prone to false positives. He is exploring a commercial path while anticipating months of work and adjustments in response to potential detectors from Anthropic.

Meyer considers a business venture amid legal uncertainties and months of work

Guillaume Meyer is contemplating transforming Watermarks Remover into a commercial activity, but he expresses uncertainty about the legality of such a product, while believing it should be feasible. He considers the tool still technical and difficult to use, viewing its simplification as a priority project. He feels it is natural for the project to evolve into a business, though he is unsure how or when, and he anticipates months of development. He also expects adjustments when Anthropic releases detectors, asserting that there is still much work to be done.

Massive virality and global contributions shape the project

After a second message on X on August 11, Guillaume Meyer reports surpassing 2 million impressions, with widespread sharing on LinkedIn and other networks. He mentions creating new accounts to follow discussions and notes that contributors from around the world mobilized after intense initial days. According to him, 99.9% of feedback is positive, while 0.1% expresses opposition. He believes the project has moved from confidentiality to becoming a public conversation topic, revealing opinions on AI watermarking. In response to presentations of the tool as "perfect," Meyer counters that the project is not finished and will not be a one-off effort.

The tool iterates on text and pixels to disrupt watermarks

Guillaume Meyer describes text watermarks based on statistical models of word choice. His approach involves checking for the presence of a watermark, generating slight variations that preserve meaning, and then repeating the process until the signal is no longer detected. He applies a similar logic to images by rewriting pixels. He claims to have released a first version on GitHub in about five hours and has been working almost full-time since, relying on previous research and experience gained with open-source models and result aggregation.

Trigger: Anthropic's announcement and the shift to open-source

The announcement of an invisible AI watermark by Anthropic served as the starting point for the project. Guillaume Meyer states that he researched these techniques, noting that other providers, such as Gemini, are already using them. He explains that he had the idea to test whether it was possible to remove such a marking and then created Watermarks Remover as an open-source tool. He presents himself as an entrepreneur with over twenty years of experience in technology.

Meyer challenges statistical approaches and claims to support attribution

Guillaume Meyer asserts that the method chosen by Anthropic relies on statistical analysis that is likely to produce false positives, which he believes would undermine trust. He cites the case where a ten-page article could be labeled as AI-generated due to a simple line modified at the end of the document. He expresses support for content attribution while rejecting the proposed technical solution. For illustration, he explains that he is not a native English speaker and uses Grammarly, which he describes as AI-powered, arguing that if an invisible watermark were added, his texts would automatically be considered AI-generated. He finally clarifies that his project is not intended to enable identity theft or content theft, but to be used on his own productions and for educational purposes.