Office workers using computers to generate and verify content with AI assistance. — Digital workplace professionals utilizing secure AI technologies like Claude in their daily workflows.

Anthropic Introduces Undetectable Text Watermarking to Claude Models to Comply with EU AI Act

Anthropic is introducing cryptographic text watermarking to future Claude models to comply with the EU AI Act’s transparency requirements. This technology, based on Google DeepMind’s SynthID-Text approach, subtly alters the mathematical randomness of word selection without altering output quality, generation speed, or pricing. Crucially, the system protects user privacy by containing zero traceable personal or organizational data.

RMN Digital Corporate Desk
New Delhi | August 20, 2026

Compliance and the Global Rollout

To comply with the European Union (EU) AI Act, Anthropic has announced that future Claude models will generate text containing an invisible watermark. This technology helps determine the likelihood that Claude was involved in producing a piece of text.

As of August 2, the EU requires AI system providers serving its market to implement methods of “marking” AI-generated content. In response, Anthropic—along with approximately 190 other signatories—signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026. Although driven by EU regulations, Anthropic is applying this text watermarking globally at launch because a durable method to restrict the feature by geographic region does not yet exist. The company is also working to roll out watermarking to older Claude models launched before August 2, 2026, over the coming months.

The Science Behind the Watermark

Large language models generate text sequentially, choosing one word at a time from a list of candidate words based on probability. When multiple words are equally sensible—such as choosing between “overcast” or “grey” to complete a sentence—the system typically relies on a standard random number generator to make the final selection.

Claude’s watermark utilizes a version of the SynthID-Text approach (originally developed and published in Nature by Google DeepMind in 2024). Instead of using an arbitrary random number generator, the watermark uses a specific cryptographic key combined with the preceding words to shift the source of randomness. This leaves a signature pattern in the generated text. While the resulting word sequences remain completely natural and indistinguishable to human readers, anyone holding the corresponding detection key can mathematically check the sequence to assign a probability that the text was generated by Claude.

Zero Impact on Performance, Cost, and Quality

A primary concern for digital professionals is whether watermarking degrades AI performance. Anthropic’s internal evaluations, as well as Google DeepMind’s live A/B traffic tests on Gemini, confirm no statistically significant differences in thumbs-up/thumbs-down ratings, creativity, or readability between watermarked and unwatermarked outputs.

Furthermore, because the watermark only shifts word-selection randomness rather than inserting hidden characters or extra words, it requires no additional tokens. As a result, watermarking has a negligible impact on model speed and introduces no extra costs for users.

Strict Privacy Guardrails

Enterprise security remains uncompromised. The watermarking key is applied broadly at the model level and carries absolutely no identifying information. It cannot be traced back to any specific user, organization, prompt, or individual chat history. The watermark does not alter content ownership or user rights under Anthropic’s commercial terms; it simply acts as a marker of Claude’s mathematical involvement.

Technical Limitations: Factual Text and Code

The efficacy of text watermarking relies on the model having choices. Consequently, it has notable technical limitations:

Factual Text: In highly rigid passages where only one specific word is correct (e.g., completing “Isaac Newton’s most famous work was called Principia…” with “Mathematica“), the watermark cannot be applied because there are no alternative word options.
Computer Code: Because code must be syntactically precise to function, it generally features far less watermarking. However, watermarks can still be applied in areas of arbitrary choice, such as code comments.
Light Editing and Proofreading: If Claude is used to lightly edit or proofread human-written text, the resulting output consists primarily of the human’s original words, leaving almost no room for a watermark to attach.
Sample Size: Detecting a watermark requires a substantial text sample to establish mathematical confidence; it does not work reliably on short passages.

Future Verification Tools and Multimedia Metadata

To help platforms and developers verify content origins, Anthropic is currently developing a watermark detection API.

For non-text media produced or processed by Claude, such as PNG, JPG, or SVG images, Anthropic is employing the C2PA (Coalition for Content Provenance and Authenticity) open industry standard. Rather than modifying the visual pixels, Claude embeds a cryptographically signed credential directly into the file’s metadata. This metadata declares Claude’s involvement and can be read by any C2PA-aware tool or via an upcoming verification tool provided by Anthropic.