New Free Tool Removes Claude Watermark a Day After Anthropic Announcement

Just days following Anthropic's announcement regarding Claude's invisible watermarking system, developers have unveiled an open-source tool intended to eliminate or diminish such AI provenance markers. The GitHub project, named watermarks-remover, addresses multiple types of AI markers present in text and files, encompassing systems linked to Claude, Google's Gemini, SynthID, OpenAI-related data, and open-source language models' provenance formats.

Anthropic disclosed on August 14 that upcoming Claude models would incorporate an imperceptible statistical watermark, as a compliance measure under the EU AI Act. However, Claude does not embed invisible Unicode characters into its text. Rather, Anthropic subtly modifies the word selection process during text generation. This resulting pattern can subsequently be identified using Anthropic's watermarking key.

Importantly, the watermark remains undetectable to human readers and does not disclose any information about the individual user, organization, or conversation.

Developers swiftly developed the watermarks-remover tool, employing various strategies contingent on the type of provenance marker involved. The first layer aims to directly eradicate elements like invisible Unicode characters, unconventional spaces, bidirectional text characters, and other concealed text markers. Claude's statistical watermark, however, necessitates a distinct approach.

The second layer endeavors to weaken statistical text watermarks by substantially rewriting the original text, altering the word choices to disrupt the underlying pattern. The project also strips provenance metadata from formats such as PNG, JPEG, WebP, PDF, DOCX, HTML, Markdown, video, and audio files.

It is crucial to note that the tool does not assure the complete removal of Claude's watermark. The developer explicitly describes statistical watermark removal as a best-effort endeavor, as Anthropic has yet to release its public watermark detection API or the keys required for independently confirming whether rewritten text still complies with its official detection system. Consequently, the tool is unable to guarantee the elimination of Claude's watermark.

The developer further warns that removing a statistical watermark demands rewriting a significant portion of the original text. Altering headings, rearranging paragraphs, or making minor edits is improbable to eradicate the signal. Such extensive rewriting can also modify the tone, precision, and overall quality of the original output.

Anthropic acknowledges that its watermark is not entirely impervious to removal. The company asserts that minor edits will likely leave enough of the signal intact for detection. Nevertheless, a comprehensive rewrite, wherein almost every word is altered, can effectively eliminate the watermark. In such instances, it becomes less apparent whether the rewritten material should still be classified as AI-generated text originating from the original model.

Anthropic is currently in the process of developing a watermark detection API that will enable users to ascertain whether Claude was likely involved in generating or processing a specific piece of text.