Brief IA
›

Anthropic Launches API to Verify Claude's Watermark

Regulation & Ethics·Tom Levy·

Anthropic Launches API to Verify Claude's Watermark

Anthropic Launches API to Verify Claude's Watermark
Key Takeaways
1Anthropic launches an API that allows for the detection of watermarks in texts generated by Claude
2Access is restricted to approved entities in the EU and companies with compliance obligations
3The method relies on SynthID, with guarantees from Anthropic being questioned by some regarding quality and transparency
Why it matters : Watermarking is becoming a regulatory requirement in Europe and raises issues of reliability and contractual use for texts generated by AI.

Anthropic is proposing an API that allows authorized organizations to verify the presence of a watermark in texts generated by Claude. The European watermarking requirement will apply to new models starting August 2, 2025. The solution relies on SynthID and comes with assurances regarding the absence of impact on quality, which have been contested by some observers.

Concerns About Text Quality and Contractual Transparency

Critics argue that the quality of texts generated by Claude could degrade if the choice of synonyms is guided by a watermark key rather than by meaning. Transparency risks related to AI fingerprint detection could pose issues in contracts prohibiting the use of AI or during fee negotiations. However, Artificial Lawyer anticipates that watermarks will often remain harmless.

Access Restricted to EU Actors and Compliance-Seeking Companies

Regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, civil society groups in the European Union, as well as companies needing to control watermarking for their own compliance can request access to the API. Anthropic states that access will gradually be opened to more users. This initiative responds to the European AI Act, which comes into effect on August 2, 2025, and mandates the integration of invisible watermarks in new models of Claude.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

A Method Derived from SynthID and Announced Guarantees

The watermarking technology employed by Anthropic uses the SynthID text technique designed by Google. It alters the way words are randomly chosen to generate a pattern recognizable through statistical analysis, a pattern that, according to Anthropic, could remain present even after certain transformations of the text. Anthropic claims that this process would offer greater reliability than traditional detectors like Pangram, while ensuring that the watermark contains no user information and has no impact on the quality or content generated. The verification API is now accessible to approved organizations.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.