01
Score candidates
The model produces a probability distribution for the next token.
context→candidate scores
Watermark / text
a conceptual field guide
The mark lives in the choices.
Anthropic’s text watermarking approach is designed to leave a statistical fingerprint without adding visible characters or changing what the text means.
Important: the lab below is an illustrative model of the idea, not Anthropic’s private implementation or a live detector.
Enter the token lab
The trick happens in low-stakes moments, when more than one next word would make sense.
01 / SCORE
What could come next?
The model scores candidate continuations before selecting one.
Theweathertodaywascoldand…
Repeated across a response
Run the steps to reveal where the signal accumulates.
One loop. Three moves.
The signal is not a visible stamp. It is a pattern in repeated, plausible choices.
02
Settle the choice
Ordinary sampling uses a random source. Keyed sampling uses the key and recent context to settle the random choice.
key+context→choice
03
Repeat, then test
Across enough flexible choices, the sequence can become statistically recognisable to someone with the key.
many choices→signal
A detector asks: is this pattern unlikely by chance?
One word proves very little. The sequence is where the evidence lives.
For each observed token, a detector can check whether it was consistent with the keyed choice rule given the preceding context.
It then aggregates the matches. Longer passages usually provide more evidence; short, exact, or heavily edited passages provide less.
Illustrative signal strength
68%little evidencemore evidence
This demo passage has enough repeated choices to look consistent with the key.
What the mark can — and can’t — say
Detection is a signal about model involvement, not a universal authorship oracle.
A detected mark means Claude was likely involved
It does not establish full authorship. Claude may have written, edited, translated, summarised, or transformed text that began elsewhere.
No mark does not prove “human-written”
A passage may be too short, heavily changed, produced by an older unsupported model, or generated by another system with a different method.
Exact facts and code offer fewer places to mark
When one continuation is clearly correct, nudging the choice would risk accuracy. Flexible prose creates more low-stakes decisions.
The mark is not a user identity tag
The watermark is intended to indicate model involvement. It is not designed to reveal a person, organisation, prompt, or chat.
Source check · 14 Aug 2026
This explainer is adapted from the supplied diagrams and checked against Anthropic’s official explainer, which says future Claude models will generate text with a watermark and that “nothing is added to the text and there are no hidden characters”.
Conceptual education only. It is not a watermark detector, authorship judgement, or statement of Anthropic’s private implementation details.