Guillaume Meyer’s open‑source tool to strip Anthropic’s AI watermark garners over 14,000 GitHub stars

Background on Anthropic’s watermark initiative

Anthropic announced that Claude models launched in the European Union on or after August 2, 2026 would embed machine‑readable watermarks from day one.

The policy states that the marks apply globally across Claude, Claude Code, Claude Cowork, Claude Tag, the company’s API and supported cloud platforms.

Two watermarking systems are used: an imperceptible text watermark embedded in model output, and signed provenance metadata for files such as PNG, JPEG and SVG based on the C2PA standard.

Anthropic claims the text watermark persists when content is copied and pasted and may survive limited editing.

However, the detection mechanism for the text watermark has not been published, leaving external verification impossible.

Meyer’s rapid development of Watermarks Remover

Guillaume Meyer, an entrepreneur and CTO who splits his time between Paris and Los Angeles, built the first version of Watermarks Remover in roughly five hours after studying Anthropic’s marking system.

His open‑source project quickly attracted more than 14,000 stars on GitHub, according to Business Insider.

By August 11, a follow‑up post about the tool generated over two million impressions on X, with additional traction on LinkedIn and other platforms.

When asked about the sudden spotlight, Meyer said, “I was not ready for the attention.”

Meyer’s GitHub profile lists him as the founder of Memo, an AI platform that evaluates advertising performance, competitor creative and customer feedback for e‑commerce brands.

He noted that two decades of technology experience and prior work with open‑source models at another startup helped him assemble the remover swiftly.

The tool initially probed the statistical signal of the text watermark, creating meaning‑preserving variations until the signal weakened.

For images, Meyer employed a comparable approach that subtly altered pixel values to diminish provenance metadata.

Since the first experiment, Watermarks Remover has expanded to strip invisible Unicode characters, delete C2PA, EXIF, XMP and other document metadata across image, document, web and media formats.

The repository also offers optional rewriting and image‑regeneration components aimed at attacking statistical or pixel‑level marks.

Implications and limitations

The project’s popularity demonstrates how quickly open‑source developers can respond to provenance controls introduced by frontier AI labs.

Anthropic concedes that its marks cannot guarantee authorship and may disappear after heavy editing.

Because Anthropic has not released its detection algorithm or detailed technical guidance, the community cannot fully verify whether Watermarks Remover defeats the text watermark.

The repository clearly distinguishes deterministic actions—such as Unicode and metadata removal—from best‑effort statistical rewriting, acknowledging that it cannot certify failure of any vendor detector.

Meyer stresses that the code is intended for educational purposes and for users working with their own content, not for impersonation or theft.

Nevertheless, the tool highlights a gap between proprietary watermark claims and the ability of independent developers to test and potentially bypass them.

As AI‑generated content becomes more prevalent, the tension between provenance enforcement and open‑source scrutiny is likely to intensify.

Stakeholders will need to monitor whether Anthropic releases a public detector or refines its watermarking techniques in response to community tools like Meyer’s.

Why This Matters is that open‑source developers can rapidly create tools that test and potentially bypass AI provenance mechanisms, as demonstrated by Meyer’s Watermarks Remover.