Claude now watermarks its text. Here's what that changes.
Anthropic announced this week that text generated by its Claude models now carries an invisible watermark. The change applies to models released after August 2, 2026, the day the EU AI Act's transparency rules took effect, and it applies worldwide, across the API, the Claude apps, Claude Code, and the rest of the product line, not only in the EU. Older models are being retrofitted. Images and other supported files get signed provenance metadata under the C2PA standard; text gets something stranger, and that something is worth understanding precisely.
What the watermark is
It is not a label, a footer, or hidden metadata. Anthropic says the mark is part of the text itself: it travels when the text is copied and pasted, and "may persist through some editing." When this guide first ran, Anthropic had not published the technical scheme and analysts expected a statistical method in the family academia has studied since 2023. On August 14 Anthropic published a FAQ confirming it: the scheme is a version of Google DeepMind's SynthID-Text. The model prefers certain word choices over statistically equivalent ones, following a secret key, and a detector can later test whether a passage followed that preference. No single word carries the mark. The signal lives in the pattern, which is why reported analyses put the minimum detectable length at very roughly a hundred tokens, why short texts carry no reliable signal, and why heavy rewriting or translation destroys it.
What the mark proves, and what it doesn't
A detected watermark means a Claude model produced or reworked that wording. It does not mean Claude wrote the ideas. People use Claude to proofread, translate, tighten, and reformat text they wrote themselves, and the output carries the mark anyway. The reverse is also true: absence of a mark proves nothing, since older models, other companies' models, short passages, and paraphrased text all come up empty. Anthropic's own framing concedes both directions, and the trade press has been blunt about it: the mark shows processing, not authorship.
What this means for our instruction files
Nothing about the skill and the portable instructions changes, and it's worth being exact about why. Our files change how text reads to a human: the vocabulary, the constructions, the punctuation habits that make a reader hear the robot. A watermark is aimed at machines, sits below anything a reader can perceive, and is applied by the model at generation time. An instruction file cannot add it, remove it, or switch it off; no custom instruction can. Since launch, this site has said that avoiding the vernacular does not defeat detectors, and that we never claim otherwise. That honesty note used to be a caveat about classifier accuracy. As of this month it is simply how the architecture works.
One related note: our checker calls no model at all. It runs entirely in your browser, so checking a draft neither adds a mark nor removes one. It just shows you what your reader already notices.
What to worry about, and what not to
Don't worry that a watermark will brand you a fraud for using a spellchecker. On the reported design, a mark cannot distinguish "Claude wrote this essay" from "Claude fixed my commas," and any teacher, editor, or employer treating it as proof of authorship is misreading it. The historical record here cuts against overconfidence in the other direction too: earlier detector tools, before watermarking, were reported to falsely flag over half of essays by non-native English speakers in one study, and several universities switched detectors off over false positives.
Do assume that where AI involvement matters, it may now be checkable. If Claude drafts or reworks your text, treat the result as carrying a durable trace of that fact, one that survives copy and paste and some light editing. In classrooms, journals, courtrooms, and workplaces with disclosure rules, the right move is the one we have always recommended: disclose the tool, don't disguise it. Trying to scrub a watermark by paraphrasing is exactly the arms race the site refuses to arm you for.
And keep some skepticism in reserve, because the biggest pieces are still missing. Anthropic has said it will publish detection details and an engineer has said a detection API is coming, but as of this update neither has shipped, and the FAQ cites quality data for the watermark without publishing detection error rates. Until that documentation lands, nobody outside Anthropic can verify a mark, which means nobody should be making accusations based on one either. When the details arrive, the trend board and this column will report what they actually say.
Sources: TechCrunch, Aug 11 2026 · The Decoder · TechTimes · explainx.ai analysis · Euronews · Anthropic FAQ, Aug 14 2026 · TechCrunch, Aug 15 2026. Updated September 15, 2026 after Anthropic published its FAQ; the analyst inferences above were confirmed by it.