Anthropic 推出水印检测 API，使第三方能够检测 Claude 的 AI 文本

Anthropic will soon offer a watermark detection API that lets third-party developers plug AI text detection into their own apps. The company uses a variant of the SynthID Text method that Google Deepmind published in Nature in 2024 for its Claude AI model. The watermark tweaks the randomness source during word selection, creating a traceable pattern. According to Anthropic, this has no effect "on the content, level of creativity, or readability of Claude's text." Google Deepmind visualized the process in 2024 with the following animation. The watermark works less reliably on short texts or fact-heavy passages where there aren't many alternative phrasings. The same goes for code. Pure corrections, where a human chose every word, won't carry the watermark either. Translations are a different story, since Claude picks all the words there. Heavy rewriting can strip the watermark out, Anthropic says in a published FAQ . Ad The watermark can only flag that Claude was likely involved in creating a text. It can't tell whether Claude wrote the whole thing or just edited it heavily, and it can't determine whether a text came from a human or a different AI model. Ad Anthropic is adding watermarking to comply with the EU AI Act . Along with roughly 190 other signatories, the company signed the EU Code of Practice on transparency for AI-generated content in July 2026. Since there's currently no technical way to limit the feature by region, watermarking launches worldwide. According to Anthropic , the company is exploring more options on an ongoing basis, with updates to follow. All models released after August 2, 2025 support watermarking out of the box. Older Claude models will get the feature in the coming months. For files, Anthropic uses the open C2PA standard, which attaches metadata without altering the file itself. Ad Anthropic's watermarking works differently from external AI detection tools like Pangram . Those services don't have access to Anthropic's keys. Instead, they scan for typical patterns in AI text, things like telltale phrasing or overused words. Checking for a watermark is a fundamentally different approach and should prove more reliable in practice. Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.