Anthropic is rolling out a way to tag Claude-generated writing so it’s easier to spot — but don’t expect a neon sticker or a secret signature visible to your eyes. To meet new rules and to be slightly less mysterious, Anthropic will add an invisible pattern into the way Claude picks words. Regular readers won’t see anything different, but tools with the right key can detect the pattern and say, with some probability, that Claude likely had a hand in the text.

How the watermark actually works

This is not about stuffing hidden characters into the output or appending “Generated by Claude” to the end of a paragraph. Instead, Anthropic alters the model’s internal randomness during generation — a tweak inspired by research approaches like SynthID-Text. When the model has several reasonable options for the next word, the watermarking process biases those random choices in a subtle, repeatable way using a secret key and a bit of the preceding context. The result is a statistical fingerprint spread across many choices that’s invisible to a reader but detectable by someone who has the key.

Key takeaways on implementation:

  • No extra characters are added to the text.
  • No additional tokens are required and generation speed is essentially unchanged.
  • The watermark is applied at generation time by nudging sampling decisions, not by post-editing the output.
  • Anthropic plans to enable watermarking globally at launch and will add it to older models still in transition over the coming months.

Limitations, exceptions, and detection

Watermarking isn’t perfect, and Anthropic is upfront about the limits. If the model has only one correct next token (think simple math or an exact code token), the watermark doesn’t nudge the choice — because changing that would be wrong or break things. That means factual answers and many code outputs carry little to no watermarking, though comments or non-essential bits of code might still show the pattern.

Other things to know:

  • Detection works best on longer, high-entropy text where many small choices create a clearer signal. Tiny snippets or heavily human-edited text may not have enough evidence to detect.
  • Translations produced entirely by Claude should show a watermark because each output word was selected by the model.
  • Anthropic will offer a detection API that estimates how likely it is Claude was involved; it’s a probability score, not a legal proof of authorship. The detector can’t say whether another model wrote it, nor can it distinguish between “Claude wrote this” and “Claude heavily edited this.”
  • For image files like PNG, JPG, and SVG, Anthropic will attach cryptographically signed provenance metadata instead of modifying the pixels themselves.

So yes — it’s clever, quiet, and statistical. The watermark aims to help systems and auditors spot AI involvement while leaving everyday reading and code behavior unchanged. If you’re wondering whether a lightly proofread or slightly rewritten passage will lose the mark: usually it won’t be fully erased unless someone thoroughly rewrites every word.