Claude Will Start Flagging Any Text It Touches — Even Text a Human Edited

Anthropic has announced that it will soon begin watermarking content that has passed through its models — and not just content that was generated, but content that was simply processed. In a help-center article, the company explained that it is introducing machine-readable marks in order to meet the requirements of the European AI Act. The law obliges providers of AI systems to label audio, images, text and video created or modified by a neural network. The rule applies to any model released after August 2, while earlier ones are granted a reprieve until December 2026.

Anthropic said that all new models worldwide, not just in the EU, will mark content “from day one.” Invisible watermarks will be embedded in text, while other files will receive a digital signature carrying provenance metadata — wherever that is supported. For images and video, the C2PA standard is used.

A Mark on Everything

Interestingly, the company has opted for a “nuke it from orbit” approach: the mark is applied to all processed content, even though the EU law explicitly exempts cases of “assistive editing” — correcting grammar, for instance — and edits that do not change the meaning. The problem is that the watermark is applied at the model level and does not distinguish full text generation from a corrected comma. As a result, Claude risks marking precisely what the law deliberately left alone: proofreading, translation, a brief summary, or the conversion of someone else's document. Anthropic acknowledges this: the original idea or text may belong to another author, but Claude's mark will remain all the same.

Easy to Bypass, Easy to Misread

Text watermarks work by means of a barely perceptible bias in word choice, distributed across the entire document and discernible only by a special tool in the aggregate. The price of this “invisibility” is that the model sometimes prefers a slightly worse word in order to preserve the signal. The mark survives copying and some edits, but disappears if you run the text through another chatbot, take a screenshot, or scrub the metadata. And as soon as Anthropic publishes a way to detect the marks — which is required for technical support under the law — building a tool to remove them will become trivial.

No less troubling is the ambiguity of the mark itself. According to the company, a detected mark means only that the content “may have been processed by Claude,” and is not conclusive proof; it can appear even on text that Claude did not create. At the same time, the absence of a mark does not guarantee that the text was not AI-generated. It is not hard to imagine a teacher who sees such a mark on a student's work and jumps to a conclusion, when all the mark actually says is “AI touched this.”

The situation is further muddled by another provision of the law, Article 50(4): fully generated text in most cases does not require a visible notice for the reader if material on a matter of public interest has undergone editorial review. Hence the paradox: at the model level, “mark everything,” but at the publication level, “you needn't mark it if a responsible editor has looked the text over.”

Failure to comply with the AI Act carries a fine of up to €15 million or 3% of global annual turnover. Anthropic promises to keep refining its watermarks and the methods for detecting them.

Source: Ars Technica