Anthropic has announced a new watermark detection API that enables third-party developers to integrate AI text detection into their applications. The API uses a modified version of the SynthID Text method, originally developed by Google Deepmind and published in Nature in 2024. According to the company, the watermark does not affect the content, creativity, or readability of Claude’s output. The feature is intended to comply with the EU AI Act and support transparency for AI-generated content, as part of the EU Code of Practice on transparency for AI-generated content, which was signed in July 2026.
The watermark works by altering the randomness source during word selection, creating a traceable pattern. However, it is less effective on short texts, fact-heavy passages, or code, where there are fewer alternative phrasings. Anthropic also noted that pure corrections, where a human selects every word, will not carry the watermark. Translations, on the other hand, are more likely to retain the watermark since Claude generates all the words. Heavy rewriting can remove the watermark, according to a published FAQ.
Anthropic’s approach differs from external AI detection tools like Pangram, which scan for typical AI text patterns rather than using a watermark. The company is exploring additional options for the feature and plans to roll it out to all models released after August 2, 2025, with older models receiving the update in the coming months. For files, the company uses the open C2PA standard to attach metadata without altering the file itself.
Source: thedecoder
