Anthropic opens watermark detection API for Claude texts

Anthropic is launching a watermark detection API that will let third-party developers verify whether text came from Claude. The company embeds an invisible statistical signature during generation, then exposes detection through an API. Developers can plug this into their own apps without needing access to Anthropic's internal keys.

The move puts Anthropic ahead of OpenAI and most other foundation-model vendors on AI-text provenance. It also marks the first commercial deployment of Google DeepMind's SynthID Text method outside Google itself.

How the watermark works

The technique manipulates the randomness source Claude uses when selecting tokens. Instead of pure randomness, the model samples from a distribution that encodes a traceable pattern. According to Anthropic, this has "no effect on the content, level of creativity, or readability of Claude's text."

Google DeepMind published the underlying SynthID Text research in Nature in 2024. Anthropic has adapted the method for Claude's architecture.

Detection is probabilistic. The API returns a likelihood score, not a binary yes or no. Short passages, code, and fact-heavy text where Claude has fewer word choices all reduce reliability. Heavy human rewriting can strip the watermark entirely. And crucially, the system cannot distinguish whether Claude wrote an entire document or just edited portions of it.

Why this is happening now

The EU AI Act. In July 2026, Anthropic joined roughly 190 signatories on the EU Code of Practice for transparency in AI-generated content. Watermarking is one of the compliance mechanisms the regulation contemplates.

Because Anthropic has no technical way to limit the feature by geography, watermarking rolls out globally. The company says it is "exploring more options on an ongoing basis, with updates to follow."

What the detection API doesn't do

The API tells you Claude was "likely involved." It cannot tell you how much. If a user pastes a paragraph, asks Claude to fix grammar, and publishes the result, the watermark may survive even though every word was the human's choice. Conversely, if a user rewrites Claude's draft heavily, the signature may vanish.

It also cannot detect text from other models. A document written by GPT-4 or Gemini will simply return low or no confidence. This is fundamentally different from external detection services like Pangram, which scan for stylistic tells across any AI output. Those tools work on pattern-matching. Anthropic's approach is cryptographic. Each has trade-offs: pattern-matching is model-agnostic but easier to evade; watermarking is harder to fool but only catches your own model's output.

For files like images and PDFs, Anthropic uses the open C2PA standard, which attaches metadata without altering the file. Watermarking and C2PA serve different purposes: one is forensic, the other is provenance. Both land together.

What's still missing

Anthropic has not disclosed API pricing, rate limits, or latency benchmarks for detection calls. It also has not said whether watermarking can be disabled for enterprise customers who want to avoid the overhead.

The bigger gap is interoperability. Google has SynthID for Gemini. Anthropic now has this. OpenAI has hinted at watermarking research but shipped nothing public. Until vendors agree on a shared detection standard, anyone building a general-purpose AI detector will still need to call multiple APIs or fall back to pattern-matching.

Source: The Decoder / Matthias Bastian