Anthropic’s Claude watermark plan sparks a fast-moving market for removal tools

Anthropic’s plan to mark AI-generated text from Claude has quickly produced a counter-movement: tools that try to remove those marks. In an August 23 Business Insider interview, Paris-based entrepreneur Guillaume Meyer described how he built an open-source project called Watermarks Remover shortly after Anthropic announced its invisible text watermarking approach. Meyer said the project spread rapidly after he posted about it, drawing millions of social impressions and attention from developers who worry that watermarking could mislabel human work touched by AI.

The dispute began with Anthropic’s decision to add machine-readable marking to supported Claude outputs as part of its commitments under the European Union’s AI Act transparency rules. Anthropic says the mark is not a visible label, does not add hidden characters, does not identify a user or organization, and should not change the meaning, readability or quality of Claude’s responses. The company has also said it plans to provide a detection API so users and third parties can check for Claude’s marks.

The technical idea behind a text watermark is statistical. Rather than attaching a tag to a document, the model slightly influences some low-stakes word choices so that the overall pattern can later be detected with the right key. That makes the mark harder to notice in normal reading and more portable when text is copied. It also makes the public debate harder, because users cannot easily see when the mark is present or judge whether a detector would treat light editing, translation or proofreading the same way as full generation.

Meyer’s criticism centers on that ambiguity. In the Business Insider interview, he argued that statistical watermarking can create false positives and “human victims,” especially in academic or professional settings where authors may use AI to edit a small part of otherwise human-written work. He said his tool is not perfect and will need to evolve as detection systems appear, but he framed the project as a way to test and challenge an approach he considers too blunt. Earlier Business Insider reporting said watermark-removal interest had risen sharply and that similar tools were appearing around the same controversy.

For Anthropic, the pressure comes from the opposite direction. Regulators want more transparency around synthetic content, and AI companies need systems that can travel with text after it leaves their products. If a label sits only inside an app interface, it disappears the moment text is copied into an email, document, website or social post. Watermarking promises a more durable signal. But durability is also why users worry. A mark that follows text can become reputationally significant even when it was created by a minor editing step.

The larger issue is that authorship is no longer binary. A document may be drafted by a person, cleaned by a model, translated by another system, fact-checked by an editor and formatted by an agent. A single watermark signal can show that a model touched the text, but it cannot by itself explain how much creative responsibility belonged to the human author. Schools, employers, publishers and courts will need to avoid treating detection results as simple proof of misconduct.

The speed of the backlash shows how difficult AI transparency will be in practice. Users want provenance and protection from synthetic spam, but they also want control over their own writing and protection from false accusation. Anthropic’s watermark may become a serious compliance tool, or it may become one move in a long arms race between markers and removers. Either way, the debate has moved from theory to software, and the next test will be whether detection APIs, policy rules and user expectations can handle the messy middle between human writing and AI assistance.