The Claude text watermark, announced by Anthropic as part of its commitments under the European Union AI Act, will be embedded in all output produced by new Claude models from 2 August 2026, with older models to follow. The marking travels with copied and pasted text and, Anthropic says, may persist through some editing.
The timing matters for a publishing industry that has been shaken by several high-profile disputes over AI-generated writing. Last month, a book agent withdrew support for the crime novel Call Me, I’ll Hide the Body after concerns were raised about its author possibly using AI. Fourteen publishers had bid for the rights and a deal had already closed before the agent apologised. The writer, Jerry Falade, denied using AI. Earlier this year, Hachette pulled Mia Ballard’s horror novel Shy Girl after similar allegations; Ballard told The New York Times she had not used AI, adding that a freelance editor had introduced AI-generated material without her direct knowledge.
How the Claude Text Watermark Works
Anthropic says the system is based on a version of the SynthID-Text approach developed by Google DeepMind. Rather than inserting visible tags or metadata, it exploits low-stakes word choices: wherever two plausible words would carry the same meaning, the model picks according to a hidden pattern. The result is a detectable signal for anyone holding the corresponding key, but one that is invisible to a human reader.
According to Anthropic’s own documentation, the watermark has a negligible impact on model speed, produces no extra tokens, and leaves the model the same price to serve and use. Importantly, it carries no personal data. Anthropic states: ‘There’s nothing in the watermark, or its key, that would allow anyone to recover any information about the user, their organization, or their chats with Claude.’
For files processed by Claude rather than text it generates from scratch, Silicon Republic reports that Anthropic applies signed metadata labels, which signal that the file was processed by Claude and allow detection of whether it has since been tampered with.
The scope is broad. The marking policy covers the consumer app, the developer-facing Claude Platform (API), Claude Code, Claude Cowork, and Claude Tag, as well as versions of Claude accessed through AWS, Google Cloud, and Microsoft Foundry, according to Euronews.
Limits of Detection and the Publishing Question
The watermark is not a silver bullet. Anthropic acknowledges that heavy editing, paraphrasing, or mixing Claude’s output with other writing can make it undetectable. There is also a subtler complication: a translation produced by Claude does carry a watermark, because in that case every word is chosen by the model. Finding the mark in a document does not, however, prove that Claude originally authored it, since using Claude only to proofread or translate a human-written text can also leave a trace.
For publishers, schools, and universities trying to investigate authorship disputes, that nuance will be critical. Detection tells investigators that Claude was involved at some stage; it does not reconstruct the degree of that involvement.
Anthropic plans to provide third parties with tools to detect the watermark, which could eventually give institutions an independent verification route rather than relying solely on the lab’s own systems.
EU Regulation and the Watermark’s Legal Context
The move is tied to Anthropic’s obligations as one of approximately 200 signatories of the EU AI Act’s Code of Practice on Transparency of AI-Generated Content, a voluntary instrument whose other signatories include OpenAI, Mistral, Meta, and Microsoft.
The underlying legal obligation, Article 50(2) of the EU AI Act, was technically postponed from 2 August to 2 December 2026 for generative AI systems already on the market before 2 August, according to analysis by Gecić Law. Anthropic chose to build the marking in on the original August deadline regardless. Non-compliance with Article 50 transparency obligations can trigger fines of up to €15 million or 3% of total global annual turnover, whichever is higher.
Google DeepMind’s SynthID, the underlying technology, adjusts the probability scores of tokens generated by a large language model to embed a watermark without affecting quality, accuracy, or speed. Google announced text and video watermarking for its Gemini app in 2024, having released an image version in 2023. Anthropic is the second major AI lab to adopt text watermarking at scale.
The real test will come the first time a publisher takes a detection result to court. Whether a watermark trace constitutes admissible evidence of authorship, or merely of proximity to Claude, is a question that has not yet been settled.