Anthropic to open Claude watermark detection API
The company says the tool will let developers check whether Claude likely helped write a passage, while it stresses that the mark is invisible, privacy-preserving and imperfect against heavy rewrites.
Anthropic said it will soon offer a watermark detection API that lets developers check whether Claude likely helped generate a passage of text. The company paired that announcement with a technical explanation of how its text watermarking system works and where it has clear limits.
As reported on 11 August, Anthropic had already begun watermarking Claude text to meet EU transparency rules. The new step extends that effort by exposing a way to detect the mark rather than only embedding it.
The company said the watermark is invisible to readers and does not change the quality of Claude’s output. In Anthropic’s account, the system does not remove randomness from generation; it shifts the source of randomness used to choose the next word, leaving a pattern that can be checked with a key.
That description closely tracks Google DeepMind’s SynthID-Text method, described in Nature in 2024, which modifies next-token sampling to create a statistical signature while preserving text quality. Anthropic said its own internal tests found no effect on creativity or readability, and a controlled side-by-side study with human raters found no difference in quality. The company also said the watermark adds no slowdown and costs the same to serve and use because it produces no extra tokens.
The system is not a universal detector. Anthropic said it can only assign a probability that Claude was involved in producing a passage, and it cannot tell whether the text was written by a human or another AI model. It also said detection is weaker on short passages, factual writing, proofreading and other cases with few edits, because there are fewer word choices for the watermark to attach to.
Code is treated differently too. Anthropic said exact outputs are left unwatermarked when no alternative token would be equally good without harming quality, although comments and other arbitrary text can still carry marks. It added that translations produced by Claude are watermarked because every word is chosen by the model.
Anthropic said the watermark cannot be traced back to a specific user, organisation or chat, and it contains no information that would let anyone recover those details. The company said it is applying watermarking globally at launch because it does not yet have a durable way to limit it by region.
For files, Anthropic said supported formats such as .png, .jpg and .svg will carry C2PA content credentials in metadata instead of an embedded text watermark. The company said C2PA is an open industry standard, and that the label is a small, cryptographically signed note showing Claude processed or made the file while leaving the file itself unchanged. Anthropic also said models launched before 2 August will gain watermarking over the coming months.