Anthropic published an explanation of Claude's text watermarking on 14 August. It is worth reading for one reason above all: the mechanism is clear enough to reason about, and reasoning about it tells you exactly where it does and does not apply to your work.
The short version for developers: your code is almost entirely unaffected, and your documentation is not.
How it works
The watermark lives in word choice. When several words would be equally valid at a given point, the model normally picks between them using randomness. The watermarking system instead uses a key plus the few words that came before to settle which one to pick.
Anthropic's own analogy is substituting the digits of pi for dice rolls in a board game: the sequence still looks random to anyone who does not know what to look for, but it is reproducible if you do. The result is a statistical pattern detectable across a passage, without altering meaning or quality.
Two practical properties follow. It is imperceptible — a reader cannot see it. And it costs nothing: no extra tokens are produced and the processing overhead is negligible, so it does not affect speed or price.
Why code is the exception
Watermarking applies minimally to code, and the reason is structural rather than a policy choice. The mechanism needs multiple equally valid options to choose between. Code does not offer them — a variable name may be flexible but the syntax, the API call and the logic are not, and a model that varies them to carry a signal produces broken code.
Anthropic says the watermark may appear in comments, which is the part of a code output that behaves like prose. So a generated function is essentially unmarked; a generated block of explanatory comments may not be.
If you were worried that AI-assisted code would arrive carrying a vendor signature, this answers it. If you were hoping a watermark would let you audit how much of your codebase was machine-written, it answers that too, in the opposite direction.
What it does and does not tell you
Anthropic is unusually clear about the limits, and the limits are the most useful part of the announcement.
- Detection needs a reasonably sized sample — longer passages give more confidence, short ones give little
- It cannot confirm human authorship. No watermark means "not marked", not "written by a person"
- It cannot identify other AI systems' output — this detects Claude, not AI in general
- It cannot distinguish "Claude wrote this" from "Claude heavily edited this"
- Light editing may preserve it; a complete rewrite removes it
Coverage across output types
The behaviour differs by what Claude produced, and the differences are logical once you know the mechanism:
- Text — all Claude models watermark their responses
- Translations — watermarked, because Claude selects every word
- Proofreading and editing — limited watermarking, because most of the original words survive
- Code — minimal, as above, possibly present in comments
- Images and files — not watermarked in the file itself; content credentials go into metadata using the C2PA standard
Privacy, scope and what comes next
The watermark contains no identifying information. It cannot be traced to an individual user, an organisation or a specific conversation — it says "Claude was likely involved", and nothing more.
It is applied globally at launch. Anthropic drove this by the EU AI Act and says it does not yet have a durable way to scope it by region, so everyone gets it while the company keeps evaluating targeted approaches.
Models launched before 2 August 2026 will gain watermarking over the coming months, and Anthropic plans to offer a public watermark detection tool. That second item is the one to watch: a detector in public hands is what turns this from a compliance artefact into something that changes behaviour.