Anthropic launched an API for verifying watermarks in text from Claude
Anthropic launched an API that allows regulators, media outlets, fact-checkers and companies to verify an invisible watermark in text from Claude — a response to EU AI Act obligations in effect since August 2026. Critics point out that rewriting removes the watermark from the text and that the method may affect quality.
On 1 September 2026, Anthropic launched a watermark verification API that allows the presence of an invisible watermark to be verified in text generated by Claude models. According to the company, access is available to regulators, law enforcement authorities, media outlets, fact-checkers, independent researchers, educational organizations and civil society organizations in the EU, as well as businesses that need to verify watermarks for their own regulatory compliance; the company plans to gradually expand the range of eligible entities. The system is based on the SynthID method from Google and adjusts the randomness of word selection to produce a statistically detectable pattern which, according to Anthropic, “may persist even with some editing of the text” and is intended to be more reliable than traditional detectors such as Pangram. The company claims that the watermark contains no user data and does not affect text quality or content — critics disagree, arguing that if the model chooses synonyms according to the watermark key instead of meaning, text quality suffers. Artificial Lawyer highlights the risk that a detectable AI trace may be a problem where contracts prohibit AI content or when negotiating compensation, although it says watermarks will be harmless in most cases.
This step follows an earlier announcement: Anthropic signed the Code of Practice for the Transparency Code under EU AI Act, whose obligations took effect on 2 August 2026 with a four-month transition period for existing products. New Claude models include an invisible watermark in generated text from their release (the SynthID-Text technology developed by Google DeepMind), while existing models are being updated gradually. Files, such as images processed by Claude models, receive digitally signed metadata under the C2PA standard, which is also used by Adobe, Google and OpenAI. According to Anthropic, labeling applies globally and across products — Claude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag — not just in the EU. The watermark survives copying of the text and, according to the company, may also persist through light editing, but a complete rewrite removes it; it is weaker in generated code because the model has less freedom in word selection.
Analyses by The Conversation and The Decoder, as well as statements from Anthropic itself, all highlight the limits of the method: the watermark does not prove that a human wrote the text, does not work reliably on short text samples, does not recognize text from another AI system and can be bypassed by rewriting with another AI tool. According to sources, OpenAI has its own watermarking ready but has not released it, partly because of concerns about an unfair impact on people writing in English as a second language. In education, experts therefore warn that neither watermarks nor detection tools such as Turnitin are a reliable solution to the problem of cheating with AI, and recommend changing assessment formats towards oral tasks or tasks developed over time instead of “detection and punishment”.
Why it matters
The new API gives third parties (authorities, media outlets, companies) a practical tool to verify whether a specific text was generated by Claude models, which matters wherever proving the origin of content is important — from compliance with EU AI Act to litigation over contracts prohibiting AI and verification in the media and education. At the same time, the limitations documented by the sources still apply: the watermark can be removed by a complete rewrite, does not work on short texts and does not prove human authorship, so it cannot be used as the sole evidence.
What was added since the original report
Verified updates
-
Anthropic launched a watermark verification API for verifying watermarks; The API is available to regulators, media outlets, fact-checkers, nonprofit organizations and businesses; The system is based on the SynthID method from Google; The API allows eligible entities to verify the presence of watermarks in text from Claude
- Anthropic launched a watermark verification API for verifying watermarks
- The API is available to regulators, media outlets, fact-checkers, nonprofit organizations and businesses
- The system is based on the SynthID method from Google
- The API allows eligible entities to verify the presence of watermarks in text from Claude
-
Watermarks can be bypassed using other AI tools; Watermarks do not work on small text samples and cannot detect all forms of AI use; OpenAI has developed watermarking but has not released it because of concerns about discrimination; Watermarking does not solve the problem of cheating with AI in education; Antropic acknowledges the practical limitations of its solution
- Watermarks can be bypassed using other AI tools
- Watermarks do not work on small text samples and cannot detect all forms of AI use
- OpenAI has developed watermarking but has not released it because of concerns about discrimination
- Watermarking does not solve the problem of cheating with AI in education
- Antropic acknowledges the practical limitations of its solution
-
Uses SynthID-Text technology from Google DeepMind; Light editing of the text does not remove the watermark, but a complete rewrite does; The watermark will be weaker in code because of the requirement for functionality; Detection APIs will be provided
- Uses SynthID-Text technology from Google DeepMind
- Light editing of the text does not remove the watermark, but a complete rewrite does
- The watermark will be weaker in code because of the requirement for functionality
- Detection APIs will be provided
-
Users have criticized the measure, particularly on Reddit; Criticism focuses on the impact of the watermark on students and journalists; Critics point to the risk of exposing lawful use of Claude; Some users argue in favor of transparency because of the risks of misuse of AI content
- Users have criticized the measure, particularly on Reddit
- Criticism focuses on the impact of the watermark on students and journalists
- Critics point to the risk of exposing lawful use of Claude
- Some users argue in favor of transparency because of the risks of misuse of AI content
-
New Claude models will have watermarks from their release, while existing models are receiving them gradually during the grace period; The C2PA standard is applied to generated images, and this is compared with the approach of Adobe, Google and OpenAI; Anthropic will share watermark detection tools in the technical documentation; The four-month grace period began on 2 August 2026 for products released before EU AI Act; Watermarks will be imperceptible and will not change content quality
- New Claude models will have watermarks from their release, while existing models are receiving them gradually during the grace period
- The C2PA standard is applied to generated images, and this is compared with the approach of Adobe, Google and OpenAI
- Anthropic will share watermark detection tools in the technical documentation
- The four-month grace period began on 2 August 2026 for products released before EU AI Act
- Watermarks will be imperceptible and will not change content quality
Two audiences, two different impacts
What this means
For individuals
Users who copy text from Claude almost unchanged (students, journalists, authors) should bear in mind that a third party can now check such text using the new API — light editing usually does not remove the watermark, while a complete rewrite does.
For a business
Companies subject to EU AI Act (or with contractual clauses prohibiting AI content) now have a tool to verify whether text was generated by Claude models, which changes the dynamics of disputes over authorship and contractual compliance, but it cannot be relied on for certainty — the watermark can be removed by rewriting, and critics say the method may affect text quality.
Risks and complianceCheck the original
Event sources
confirmed by 4 independent sources · 5 publishers, 4 independent. We count feeds from the same owner only once.