Anthropic has embedded invisible, machine-readable watermarks into all AI-generated text produced by its Claude models globally to comply with the European Union’s AI Act, which came into effect in early August 2026 [1]. This watermarking modifies Claude’s choice of words and phrasing to create a detectable pattern invisible to human readers but identifiable by algorithms [1, 2]. Anthropic stated it applies the watermark worldwide because it lacks regional filtering capabilities [2].
Under the EU AI Act, model providers face fines up to 3% of their annual turnover if synthetic content is not clearly labeled [1]. The law’s Article 50(2) excludes lightly edited or proofread AI outputs from mandatory watermarking, but Anthropic’s approach goes beyond the legal minimum. It marks even lightly edited or proofread text, not just fully AI-generated content [2]. Caroline De Cock of the Lisbon Council criticized Anthropic’s broad approach, saying, "Article 50 called for a scalpel to catch deepfakes and mass synthetic generation, but Anthropic is using a sledgehammer that hits simple proofreading and translation" [2]. Milos Rusic, CEO of Deepset, explained this comes from the ability to attach a "provenance signal" to word choices even if Claude did not originate the idea or document [2]. The company said lighter editing by Claude is less likely to trigger the watermark because there is "very little (if anything) for the watermark to attach to" in such cases [2].
The watermark remains detectable even when text is copied, pasted, translated, or summarized by Claude [2]. This technical challenge is unlike watermarking in images, where embedding does not alter wording or meaning; digital text watermarking affects phrasing and vocabulary [3].
Following Anthropic’s announcement around August 17, 2026, developers quickly produced code to remove Claude’s invisible watermarks. Guillaume Meyer published such code within four hours, and it went viral on GitHub and social media. Meyer's code has over 100 contributors on GitHub and was bookmarked more than 20,000 times on X (formerly Twitter) [1]. Meyer said, "I'm not against transparency, and I'm all for content attribution. I just think watermarking in itself is a really bad solution, because it has major drawbacks and risks." He warned using the watermark as evidence could lead to unfair judgments as Anthropic itself admits it only provides a probability estimate that text was generated by Claude [1].
Users have canceled subscriptions over the watermark, and some developers see circumventing it as a challenge or opposition to mandatory AI labeling [1]. Meanwhile, Forbes published an analysis clarifying widespread confusion and misinformation about Anthropic’s watermarking, especially regarding proofreading and minor edits [3]. Business Insider detailed how Anthropic’s watermarking exceeds EU legal requirements and sparked controversy [2].
Anthropic’s watermarking aims to reduce regulatory risk and fits its safety-first branding approach [2]. The next step will be monitoring EU enforcement and responses from users and regulators as the watermark mandate is enforced.
