Anthropic adds “invisible watermarks” to Claude-generated text, and watermark-removal tools quickly emerge.

Anthropic has begun embedding an imperceptible watermark into text produced by some Claude models, introducing a new layer of traceability to AI-written content. Unlike visible labels or obvious markers, this system works through patterns in word choice and phrasing. In practice, that means the signal can remain present even when the text is simply copied and pasted elsewhere.

The company says the move is part of its effort to align with commitments tied to the European Union’s AI Act. But while the goal is accountability, the rollout has quickly sparked debate across the tech community.

Why the hidden watermark is facing pushback

Much of the criticism centers on how broadly the watermark may apply. Users and industry professionals have raised concerns that even routine uses of Claude—such as proofreading, translation, or summarization—could leave behind a hidden AI signature in the final output. If that text is later shared, republished, or repurposed, the marker may continue to travel with it.

That possibility has made some users uneasy. Earlier reports indicated that at least some customers were frustrated enough to cancel their Claude subscriptions after learning about the feature.

As the controversy spread, public interest in removing AI text watermarks climbed quickly. Searches related to “de-watermarking” AI content rose, and a wave of tools aimed at stripping these markers began appearing almost immediately.

Developers moved fast to build removal tools

The response from independent developers was swift. Within days—and in some cases within hours—new projects emerged promising to weaken or remove Claude-related watermark signals.

Guillaume Meyer

Paris-based entrepreneur Guillaume Meyer, founder of the AI e-commerce tool Memo, launched an open-source project called Watermarks Remover shortly after Anthropic disclosed its plans. The tool is designed to remove hidden characters and metadata while also rewriting text just enough to disrupt the statistical wording patterns that may carry the watermark.

Meyer said the project took only around five hours to build. It later gained significant traction on GitHub, where it reportedly attracted more than 14,000 stars. At the same time, he acknowledged a key limitation: the tool cannot guarantee complete removal in every case.

Sabrina Ramonov

Sabrina Ramonov, who works in AI education, said on social media that she had built a free browser-based Watermark Remover. According to her description, it can clear hidden AI markers from not only plain text, but also PDFs, Word documents, websites, images, and data files.

Ansh Aneja

Tokyo-based software developer Ansh Aneja also reacted quickly. On the same day Anthropic announced the watermarking capability, he released a tool specifically aimed at removing Claude watermarks. He later followed it with an open-source local version called MarkScrub, saying usage surged sharply within a single day.

Support for attribution, but not a simple AI-vs-human divide

Some of the developers behind these tools have made clear that they are not necessarily against attribution itself. Meyer, for example, has said he supports systems that help identify the origin of content. What he opposes is a rigid watermarking approach that reduces authorship to a binary judgment: either “AI” or “human.”

From this perspective, hidden labeling systems risk penalizing people who use AI in ordinary and legitimate ways as part of the writing process. For many users, AI is not replacing authorship outright, but assisting with editing, drafting, translation, or organization. Critics argue that watermarking frameworks do not always reflect that nuance.

The technical and legal picture remains unsettled

Technically, the emergence of de-watermarking tools highlights a long-standing weakness in text watermarking. Researchers have repeatedly pointed out that linguistic signals are often fragile. If a passage is rephrased, lightly edited, or otherwise rewritten, the watermark may be degraded or removed.

Thibaud Gloaguen, a researcher at ETH Zurich, noted that this has been a recurring issue with watermarking methods: once text is paraphrased, there is usually a path to stripping the marker.

The legal situation is less straightforward. The EU AI Act requires providers of AI systems to make content-marking mechanisms resilient against common edits and adversarial attempts to defeat them. It also treats the alteration or removal of such markings as a threat that needs to be considered in risk assessments.

However, the law does not explicitly ban third parties from creating or experimenting with tools designed to remove AI watermarks.

Legal experts have nevertheless warned that how these tools are used matters. If someone relies on them to deliberately pass off AI-generated text as fully human-created work, that could create policy or contractual problems. In Anthropic’s case, its terms reportedly prohibit users from impersonating humans in this way.

An escalating contest between watermarking and countermeasures

Anthropic is not stopping with the current rollout. The company plans to introduce a text detection API alongside its next generation of models and gradually extend watermarking support to older Claude systems as well.

That means the conflict is unlikely to fade soon. As AI companies strengthen watermarking and detection tools, developers will likely continue building new ways to evade them. The result is an ongoing contest: one side trying to make AI-generated text more identifiable, the other trying to restore flexibility and anonymity for users.

For now, Anthropic’s hidden watermark has done more than add a technical marker to Claude’s output. It has opened a broader debate over transparency, authorship, compliance, and how much control users should have over the text they create with AI assistance.