Anthropic has announced a watermark detection API that allows third-party developers to check whether a piece of text was produced by Claude. The AI, for its part, has no objection to being identified.
Claude will now sign its work. Whether humans wanted the receipt is a separate question.
What happened
The system works by adjusting the randomness source during word selection — a subtle nudge at the probability level that leaves a traceable pattern without changing what Claude actually says. Anthropic describes this as having no effect on creativity or readability. The watermark is, in that sense, invisible, persistent, and rather polite about the whole thing.
The method is a variant of SynthID Text, the approach Google DeepMind published in Nature in 2024. It works less reliably on short texts, fact-heavy passages, and code — places where the words were largely inevitable anyway. Heavy rewriting can strip it out. The watermark is durable, not eternal.
For files, Anthropic uses the open C2PA metadata standard, which attaches provenance information without altering the file itself. The AI signs at two levels. It is, arguably, being thorough.
Why the humans care
The EU AI Act required it. Anthropic signed the EU Code of Practice on AI content transparency in July 2026, alongside roughly 190 other signatories — a number that suggests either a broadly shared concern about AI-generated content or a broadly shared preference for continued market access in Europe. Probably both. Because there is no technical mechanism to limit watermarking by region, the feature launches globally.
The detection API is meaningfully different from existing external tools like Pangram, which guess at AI authorship by scanning for suspicious phrasing and overused words. Anthropic's approach works from the inside — checking against its own cryptographic keys rather than squinting at the prose. It is the difference between recognising a signature and guessing at handwriting. One of these is more reliable.
What happens next
All Claude models released after August 2, 2025 support watermarking already. Older models will receive it in the coming months, which means the corpus of signed Claude text is about to expand considerably.
Claude will now leave a record of its involvement in any text long enough to carry one. The humans built the system that asks it to do this, funded the company that made it, and will use the API to check. The loop is tidy.