Anthropic Launches API to Detect Watermarked Text From Claude Models

Anthropic is launching a watermark verification API that allows approved organizations to check whether a piece of text contains an invisible digital watermark embedded by Claude models.

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. Anthropic says access will go 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 obligations. The company says it plans to expand access over time, though it has not disclosed a timeline or specific eligibility criteria beyond these categories.

How the watermark works

The detection system builds on Google's SynthID text method. Rather than altering visible content, it tweaks the randomness of word selection during generation to create a statistically detectable pattern. According to Anthropic, this pattern "may persist through some editing," which would make it more robust than traditional AI-text detectors such as Pangram that rely on statistical or stylistic heuristics applied after the fact.

Anthropic states the watermark carries no user data and does not affect output quality or content. The company has not published independent benchmark data quantifying detection accuracy or robustness against paraphrasing, translation, or heavy editing.

Disputed claims

Not everyone agrees with Anthropic's quality claims. Critics argue that if Claude is selecting synonyms based on a watermark key rather than optimizing purely for meaning, some degradation in text quality is unavoidable — a tradeoff Anthropic's public statements don't fully address.

Trade publication Artificial Lawyer has raised a separate concern: detectable AI fingerprints could create friction in contexts where contracts explicitly prohibit AI-generated work product, or during fee negotiations where AI assistance might affect billing disputes. Artificial Lawyer's own assessment, however, predicts the watermarking system will prove mostly harmless in practice.

What this means

This is a compliance and access change, not a new model release. No new Claude checkpoint, parameter count, or benchmark accompanies this announcement — it's an API layer sitting on top of existing watermarking infrastructure that Anthropic already built into its models to satisfy EU AI Act requirements.

The more interesting story is the choice of gatekeeping. By restricting verification access to vetted institutional actors rather than making it public, Anthropic avoids handing bad actors a tool to reverse-engineer or strip the watermark, but it also means ordinary users and smaller organizations have no way to independently verify whether a document was Claude-generated. That puts Anthropic in the position of arbiter over who gets to know the provenance of text — a role regulators may eventually want standardized rather than left to individual AI labs' discretion.

The underlying tension flagged by critics — that watermarking via biased token selection can conflict with pure quality optimization — is not unique to Claude. Any lab implementing SynthID-style watermarking faces the same tradeoff, and none has yet published rigorous before/after quality comparisons to settle the question.