Anthropic has added a watermark to text generated by Claude.
The San Francisco‑based firm says the machine‑readable mark travels with the text when it is copied or pasted, and it will appear across all stable versions of the Claude service. The feature is also intended to curb the flood of AI‑generated content that the company says has overwhelmed online platforms.
How the watermark works
Under the EU AI Act’s Transparency Code, users must be informed when they interact directly with an AI system, and a machine‑readable mark must enable detection of AI‑generated or manipulated content. Anthropic’s feature is designed to meet that requirement. Compliance is crucial for companies that wish to sell AI services in European markets, where the Act can restrict products that lack such transparency features.
Anthropic notes that the watermark has limits; even if a mark is detected, the company cautions it may “not [be] fully conclusive” that Claude created the content.
Reactions from tech and academia
“Some people see this as a first step towards future detection tools – which, going by track record, don't work reliably,” said Tina Austin, director of AI education and learning systems at UCLA.
“Because a mark only proves Claude touched the content, not that it wrote all of it. And no mark doesn't prove human authorship either, since heavy editing or format conversion can strip it out.”
— Tina Austin, director of AI education and learning systems at UCLA
“These models were built by copying everyone's work without asking. Our writing, our books, our posts. And now they get to put a watermark on writing that was ours to begin with and quietly decide what still counts as human,”
— Saif Kidwai, founder, Alnaas Advisory
Tech giants such as OpenAI, Google, Meta Platforms, Microsoft, LinkedIn and Snap have already introduced measures to comply with similar transparency rules, showing a broader industry push toward watermarking and flagging AI‑generated content. LinkedIn and Snap, for example, have added a button that lets users flag AI‑generated material.
In academia, Tim Mousel, an AI faculty fellow at Lone Star College in Texas, sees the watermark as a possible piece of evidence when reviewing student work, while also questioning whether it signals the “beginning of the end of students using AI to write their papers.” The debate has even reached commencement ceremonies, where former Google chief executive Eric Schmidt was booed after praising AI, highlighting the polarising nature of the technology in education. Mousel cautions that while a watermark can serve as evidence, it does not alone prove misconduct, underscoring the need for broader assessment.
Overall, the move opens a new layer of scrutiny for AI‑generated text, but the lack of technical detail leaves many stakeholders waiting for clearer guidance.
Frequently asked questions
What is Anthropic watermarking and how does it work?
Anthropic watermarking embeds a machine-readable mark in Claude-generated text to satisfy EU AI Act rules, enabling detection of AI-generated or manipulated content.
Will the Anthropic watermark prove authorship conclusively?
Experts warn that the watermark may not prove authorship conclusively, as it only indicates that Claude touched the content, not that it wrote all of it.
Why is Anthropic introducing watermarking to Claude?
Anthropic is introducing watermarking to meet EU AI Act transparency rules, which are crucial for companies that wish to sell AI services in European markets.
