Anthropic has published further technical detail on how its chatbot Claude marks the text it generates, addressing questions from users about whether the system can be edited away, how it affects programming code, and what exactly it detects.
According to TechCrunch, which first reported the update, Anthropic's blog post was designed to answer three basic questions users kept raising: how the watermarking actually works, whether it can be hidden through editing, and how it affects code.
How the watermark is built into the text
Large language models like Claude generate text one word, or token, at a time, choosing from a list of statistically likely candidates at each step. Anthropic's watermarking system nudges those choices using a secret key, so that across a long enough passage the pattern of word selection becomes a detectable statistical signature, even though no individual word looks unusual.
Anthropic said it is adopting the SynthID-Text approach that Google DeepMind published in 2024, and that it plans to release a public watermark detection tool so that users and third parties can check whether a piece of text carries the mark. Because the system operates during the word-by-word generation process itself, rather than being stamped onto a finished response, Anthropic said the change should be imperceptible to a reader. As the Forbes technology desk reported, the method works by "statistically biasing Claude's word choices according to a key held by Anthropic," with individual choices looking unremarkable but a detectable pattern building up over a longer stretch of text.
“"To a reader, a watermarked response is indistinguishable from an unwatermarked one."”
Why now, and does it apply outside Europe
The change is driven by the EU's Artificial Intelligence Act, specifically Article 50 and its associated Code of Practice on transparency, which requires companies to use technical means that make it possible to identify content that has been generated or heavily edited by AI. These transparency obligations became enforceable on 2 August 2026, applying to models launched from that date onward, with older systems covered by a transition period.
Notably, Anthropic has chosen to apply the watermark worldwide rather than confining it to European users, a decision that has drawn attention because it means the requirement effectively becomes a global default rather than a regional carve-out. The mark is applied at the model level, meaning it works consistently whether someone uses Claude through the API, the claude.ai website, Claude Code, or via cloud partners such as AWS, Google Cloud or Microsoft's Azure-based Foundry service.
The company has separately confirmed it is applying signed provenance metadata, based on the Coalition for Content Provenance and Authenticity (C2PA) standard used across the media and technology industry, to image files such as PNG and JPG outputs. That system is distinct from the text watermark and works more like a digital signature attached to a file rather than a pattern embedded in the content itself.
What survives editing, and what it means for code
Anthropic has acknowledged the mark is not indestructible. Heavy rewriting, translation into another language, or very short passages of text may not carry enough of the statistical pattern to be reliably detected, and the company has been explicit that a failure to detect a watermark does not prove a text is human-written, only that no mark was found.
Programming code presents its own complications, since developers routinely reformat, rename variables in, or strip comments from code after it is generated, any of which could disrupt a watermark pattern. Anthropic addressed this directly, saying the effect on functional code should be minimal even if some of the discretionary elements, such as comments, carry the mark.
“"By definition, it will have a negligible effect on the actual code produced."”
The announcement has generated debate among Claude's user base. Reactions ranged from users on Reddit calling the move a conspiracy against ordinary users, to others arguing that objecting to transparency mainly serves people who want to misrepresent AI-assisted work as their own. Some users reportedly cancelled subscriptions over the change, according to reporting cited by TechCrunch, illustrating how contentious even a largely invisible compliance measure can become once flagged publicly.
For readers across Europe and beyond, the practical upshot is straightforward: text and code produced by newer Claude models will now carry an underlying, usually undetectable signature intended to help identify AI involvement, a small but symbolically significant step as regulators push AI companies toward greater transparency about what their systems produce.
This article is free to read. It always will be — no paywall, no account, no tracking.
