Anthropic Launches Claude Watermark Detection API

Anthropic will soon release a watermark detection API that allows third-party developers to integrate AI text detection directly into their applications. The system utilizes a modified version of Google DeepMind's SynthID Text method, which was published in Nature in 2024. By subtly adjusting the randomness source during word selection, the technology embeds a traceable pattern into Claude's outputs. Anthropic claims this process does not affect the readability, creativity, or overall quality of the generated text.

The watermarking system has several technical limitations. It is less effective on short texts, code, or highly factual passages where word choice is constrained. While heavy rewriting can remove the watermark, translations remain detectable because Claude selects the final phrasing. Furthermore, the tool only indicates whether Claude was involved in generating or heavily editing a text; it cannot distinguish between human writing and other AI models. For files, Anthropic relies on the open C2PA standard to attach metadata without modifying the file content.

This initiative helps Anthropic comply with the EU AI Act. In July 2026, the company joined roughly 190 other signatories in signing the EU Code of Practice on transparency for AI-generated content. Because regional enforcement is technically difficult, the watermarking feature will launch globally. All Claude models released after August 2, 2025, support this watermarking out of the box, while older models will receive the update in the coming months.

For developers and platform moderators, this API offers a more reliable alternative to traditional AI detectors like Pangram. Instead of guessing based on overused phrases or stylistic patterns, the API verifies the watermark directly. This shift provides practitioners with a concrete verification method, reducing false positives and helping platforms enforce content policies with greater confidence.