Anthropic opens Claude watermark detection API to regulators and media
Anthropic is launching a watermark verification API that lets approved organizations check whether a piece of text contains a digital watermark from Claude. The move responds to the EU AI Act, which since August 2, 2025 has required new Claude models to embed invisible watermarks in their text output. Access goes to regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups, as well as enterprises that need to verify watermarking for their own compliance. Anthropic plans to expand access over time, though no process or timeline for that expansion is given.
The system builds on Google's SynthID text method, tweaking the randomness Claude uses when selecting words to create a statistically detectable pattern. Anthropic says this pattern "may persist through some editing," which the article's author argues should make it far more reliable than traditional AI-text detectors like Pangram.
Anthropic states the watermark contains no user data and affects neither the quality nor the content of Claude's output. Critics disagree: if Claude is choosing synonyms based on a watermark key rather than meaning, text quality suffers as a result. Trade publication Artificial Lawyer raises a separate concern, warning that detectable AI fingerprints could become a problem in settings where contracts ban AI use or during fee negotiations, though it predicts the watermarks are mostly harmless overall.
Key facts
- Anthropic is launching a watermark verification API so approved organizations can check text for Claude's invisible digital watermark.
- The EU AI Act has required new Claude models to embed such watermarks since August 2, 2025.
- Access is offered to regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, EU civil society groups, and enterprises needing compliance verification, with Anthropic planning to expand access over time.
- The system builds on Google's SynthID text method, adjusting word-selection randomness to leave a statistically detectable pattern that Anthropic says may persist through some editing.
- Anthropic says the watermark carries no user data and does not affect quality or content, but critics argue synonym choices driven by a watermark key can hurt text quality, and Artificial Lawyer warns the fingerprints could cause friction where contracts ban AI use.
Why it matters
This is Anthropic's first mechanism for letting outside parties independently verify whether a piece of text came from Claude, and it exists because a law requires it: the EU AI Act has mandated invisible watermarking in new Claude models' text output since August 2, 2025. Turning that mandate into an API that regulators, law enforcement and civil society groups can actually query is a concrete compliance step, not just a policy statement.
Who it affects
Access is limited to approved organizations: regulators, law enforcement, media outlets, fact-checkers, independent researchers, educational organizations and EU civil society groups, plus enterprises that need to verify watermarking for their own compliance obligations. General users and the public are not named as having direct access; Anthropic says it plans to expand access over time, though no timeline is given.
How to use it
Approved organizations request access to the verification API and can then check submitted text for the presence of Claude's watermark. No pricing, tiers or application process are described in the source; the mechanism is limited to text produced by Claude models covered by the EU AI Act requirement.
How solid is it
The detection method builds on Google's SynthID text approach, adjusting the randomness Claude uses when picking words to leave a statistically detectable pattern. Anthropic says this pattern may persist through some editing, which the article's author argues should make it more reliable than older detectors such as Pangram, though that comparison is the author's assessment rather than an independently verified benchmark.
Risks and caveats
Critics dispute Anthropic's claim that the watermark leaves quality and content untouched, arguing that picking synonyms according to a watermark key rather than meaning can degrade text quality. Artificial Lawyer separately flags a transparency risk: detectable AI fingerprints could create problems in contracts that ban AI use or during fee negotiations, though it predicts the watermarks are mostly harmless in practice. Anthropic's response to these transparency concerns is not reported.
“may persist through some editing”
— Anthropic