Anthropic opens Claude AI text detection to regulators, media, fact-checkers, and others
| Source: THE DECODER
Tags: EU AI Act, Anthropic, Claude, watermarking, SynthID, content detection, regulatory compliance
Anthropic launches a watermark verification API built on Google's SynthID method, letting approved regulators, media, and fact-checkers detect whether text was written by Claude — a direct response to the EU AI Act's August 2025 mandate requiring invisible watermarks in AI-generated content.
Details
The EU AI Act, in force since August 2, 2025, requires new Claude models to embed invisible watermarks in text output. Anthropic's new verification API lets approved organizations — regulators, law enforcement, media outlets, fact-checkers, researchers, and EU civil society groups — check whether a piece of text carries Claude's digital watermark. Enterprises needing compliance verification can also apply. The technology adapts Google's SynthID approach: it subtly tweaks word-selection randomness using a cryptographic key, creating a statistically detectable pattern that Anthropic says may persist through some light editing. This is designed to be more reliable than traditional AI detectors like Pangram. Anthropic maintains the watermark contains no user data and doesn't degrade text quality. Critics dispute this: if Claude selects words based on a watermark key rather than optimal meaning, output quality is implicitly constrained. Trade publication Artificial Lawyer also flags contractual risks — detectable AI fingerprints could become legally significant in agreements that prohibit AI use or affect professional fee negotiations. Access is rolling out gradually to vetted institutions; no timeline was given for broader availability.