Anthropic is now letting outside organizations verify whether a piece of text came from Claude. The company’s watermark verification API, announced this week, gives regulators, law enforcement, media outlets, fact-checkers, independent researchers, and EU civil society groups a way to check for the invisible digital fingerprints embedded in Claude’s output.
The timing isn’t accidental. Since August 2, 2025, the EU AI Act has required new Claude models to carry invisible watermarks in their text. Anthropic says enterprises that need to prove their own compliance can also request access, and the company plans to expand the program over time.
The underlying technique borrows from Google’s SynthID approach. Rather than tagging text with obvious markers, Anthropic tweaks the randomness of word selection to create a statistical pattern that survives some editing. That’s a meaningful step up from older detectors like Pangram, which often crumble the moment someone rephrases a sentence. Anthropic insists the watermark carries no user data and doesn’t hurt output quality. Not everyone buys that. If Claude picks synonyms based on a watermark key instead of meaning, some critics argue, text quality inevitably degrades.
Trade publication Artificial Lawyer points to a different problem: detectable AI fingerprints could create headaches in contract negotiations or anywhere AI use is banned outright. Still, the publication predicts the watermarks are mostly harmless in practice. The bigger question is whether statistical watermarks that survive editing will actually hold up in the wild — or whether they’ll follow the same arc as every other AI-detection tool that promised more than it delivered.
