Guillaume Meyer's watermark remover tops 14,000 GitHub stars as Claude detection details remain unpublished

Why it matters

Watermarks Remover shows how quickly open-source developers can challenge provenance controls introduced by frontier AI labs. Anthropic concedes that its marks cannot establish authorship and may disappear after heavy editing, while the absence of a public detector prevents outside researchers from testing removal claims against Anthropic's own mechanism.

Guillaume Meyer (@guillaumemeyer) turned a few hours of curiosity about Anthropic's new AI watermark into Watermarks Remover, an open-source project that had more than 14,000 GitHub stars at the time of the Business Insider reporting.

Meyer says he built the first version in roughly five hours after researching Anthropic's marking system. By August 11, his second post about the project had generated more than 2 million impressions on X, according to his account of the launch in Business Insider. The attention also spread to LinkedIn and other social networks.

Meyer told Business Insider, "I was not ready for the attention."

The rush caught a founder who was already building another AI product. Meyer's GitHub profile describes him as an entrepreneur, adviser and CTO living between Paris and Los Angeles, and identifies him as the founder of Memo, an AI platform that analyzes advertising performance, competitor creative and customer feedback for e-commerce brands. Meyer told Business Insider he has spent more than 20 years in technology and previously worked with multiple open-source models at another startup. That experience, he said, helped him assemble the first remover quickly.

A compliance announcement became a build prompt

Anthropic created the opening when it detailed plans to mark output from supported Claude models. Under Anthropic's published policy, Claude models launched in the European Union on or after August 2, 2026 support machine-readable marking from launch. The marks apply worldwide across Claude, Claude Code, Claude Cowork, Claude Tag, Anthropic's API and supported cloud platforms.

Anthropic uses two systems. Generated text receives an imperceptible watermark embedded through the model's output, while supported files such as PNG, JPEG and SVG receive signed provenance metadata based on the C2PA standard. Anthropic says the text watermark travels when content is copied and pasted and may survive some editing.

Meyer treated that announcement as a testable engineering claim. His original approach checked for a statistical signal, produced meaning-preserving variations and repeated the process until the signal weakened. For images, he described an analogous process that alters pixels.

The tool that exists today is considerably broader than that first experiment. Watermarks Remover can inspect and strip invisible Unicode characters and remove C2PA, EXIF, XMP and document metadata across image, document, web and media formats. It also offers optional rewriting and image-regeneration components intended to attack statistical or pixel-level marks.

The verifiable parts are the less controversial ones

The project's popularity supplies no proof that it defeats Anthropic's watermark. Removing visible metadata and invisible Unicode characters can be checked directly. Statistical text marks are harder because the signal lives in word choices, and Anthropic has yet to publish the detection mechanism or full technical guidance needed for outside testing.

Meyer's documentation acknowledges that limit. The repository reports deterministic Unicode and metadata actions while describing statistical rewriting as best-effort. It cannot certify that a vendor detector will fail. Its experimental image backends likewise alter content while attempting to weaken pixel-level marks.

Meyer attributes a narrower purpose to the code. He told Business Insider that it is intended for educational use and for people working with their own content, rather than impersonation or theft. He also warned that the tool remains difficult to use and will require months of adjustments as Anthropic releases or changes its detectors.

Watermarks Remover already has company in a market forming around Claude's policy. Unmarker is a commercial, pre-launch service advertising Claude-focused scoring, rewriting, meaning checks and planned API access. Unmarkr focuses on images and claims to remove visible marks, hidden AI tags, SynthID and C2PA metadata. The Watermarks Remover repository warns that removal is not guaranteed, while Nature's reporting describes the broader limits of watermark-based attribution. Anthropic has not released its public detector.

Anthropic's documentation acknowledges the ambiguity behind Meyer's critique. A detected mark indicates that Claude may have processed content; it does not establish that Claude wrote the original material. Anthropic lists proofreading, translation, summarization and file conversion as examples where a mark could attach to work originating elsewhere. Heavy editing, paraphrasing or translation can also make a mark undetectable.

That creates an enforcement problem. A signal that survives ordinary editing can follow legitimate AI-assisted work. Extensive rewriting may erase the signal, though it can also change style, wording and precision. Anthropic's own policy says a missing mark cannot establish that material avoided AI processing.

Meyer is arguing for attribution with more detail

Meyer says he supports content attribution. His objection centers on techniques that can collapse a mixed human-and-machine workflow into a binary AI label. He points to proofreading as a common case: a writer may originate the work, use an AI system to correct a small portion and still produce marked output.

Provenance standards offer a record of processing history rather than relying solely on a statistical signal embedded in prose. A C2PA implementation guide describes Content Credentials as tamper-evident, cryptographically signed manifests that can record provenance and disclose whether material was AI-generated or AI-modified. Those credentials apply to files; they do not establish authorship by themselves.

Watermarks Remover can strip some hard-bound C2PA metadata, although its documentation says soft-bound signals that reconnect a file to a remote credential remain outside the project's scope. Metadata removal is a familiar file operation. Weakening an embedded statistical signal requires changing the content itself, which makes the result harder to measure without the watermark provider's detector.

The side project is already becoming product-shaped

Meyer is considering turning the project into a business, though he has not settled on a product or timetable. No funding round, valuation or named investors have been disclosed for the project.

The current tool remains too technical for many users, Meyer says. A commercial version would need to simplify that operation while continuing to change alongside Anthropic's detectors. Meyer has also questioned whether commercialization would be legally possible.

Anthropic announced its marking system before releasing the public detector needed to evaluate it. Meyer supplied open code for testing the surrounding assumptions.

The result is a fast feedback loop between a frontier AI lab and an independent founder. Anthropic's compliance mechanism created a new technical surface, and Meyer put a remover against it within hours. For now, the repository labels statistical rewriting best-effort, while Anthropic says details of its detection mechanism remain forthcoming.