Paste text → see how it was marked, where, and strip it. Everything runs locally in this file; nothing is uploaded.
Read this first. A statistical watermark (Google DeepMind's SynthID-Text, Anthropic's
Claude watermark — global since 2 August 2026 across the API, Claude, Claude Code, Cowork and Tag —
or the green-list logit-biasing scheme of Kirchenbauer et al. 2023) adds no characters. The
watermark is the model's choice of words: at each step a secret key splits the vocabulary into
favoured and unfavoured tokens, and the sampler is nudged toward the favoured ones. Verifying it requires
that secret key, which only the model provider holds — so this panel cannot detect it, and neither can
any offline tool. That is also why it survives copy-paste, retyping and reformatting perfectly, and
why paraphrasing is the thing that breaks it.
On Claude specifically: Anthropic says it will publish verification tooling, but has not done so
yet and has given no date. Until then nothing — including the Hidden characters tab — can check it. Note also
what a hit would and would not mean: a detected mark proves the text passed through Claude, not
that Claude wrote it (proofreading, translation and summarising all leave the mark), and its absence
proves nothing, since editing, format conversion and screenshots strip it.
What follows are style statistics. They tell you whether a real detector would even have enough
signal to work with, and they surface the surface-level habits people mistake for watermarks. Treat every
number below as description, not evidence. None of it can prove or disprove AI authorship.
Input
The sample is a genuine assistant-written passage — not a mock-up, and not hand-tuned
to trip the metrics. It was written by Claude, so if Anthropic’s watermark is active it is present
in those exact words. This panel still cannot see it. That is the whole point of this tab.
Detector viability
Style statistics — descriptive only
Recurring phrases — training-distribution habits, not watermarks
Synonym rewrite — swap words, keep every sentence intact
Substitutes whole words in place, from a synonym bank, without
touching sentence structure, order or punctuation. Inflected forms are kept in separate synsets so
gender and number agreement survives; capitalisation is preserved; code spans, URLs and numbers are
never touched.
135710
level 22 substitutions per paragraph
language: —
Changes — removed / added
1 · Input — plain text
Paste as plain text straight from the original source. Retyping it, or passing it
through an editor that normalizes Unicode, destroys the evidence before it reaches this box.
2 · How it was done
3 · Where — red markers
ZW zero-width · TAG tag block · VS variation selector ·
BDI bidi control · NB odd space · underlined lookalike letter
— hover any marker for its codepoint.
Position list
#
Ln:Col
Codepoint
Name
Kind
4 · Recovered payload
5 · Clean the text
Rich paste — the clipboard's HTML flavour
Copying from a web page or document puts two versions on the
clipboard: text/plain and text/html. Marks that live in the HTML — text hidden
by CSS, comments, custom attributes — vanish when you paste into a plain-text box, which is why the Hidden characters
tab cannot see them. Paste here to capture the HTML flavour instead.
How — findings in the markup
Text hidden from view but present in the markup
Where — hidden nodes forced visible, outlined red
Clean text extraction
Visible text only, with hidden nodes dropped and the Hidden characters
filters applied.
What the clipboard can never carry. C2PA / Content Credentials — the cryptographically signed
provenance manifest attached to AI-generated files — lives in the file container, not in the text.
Selecting text and copying it strips it instantly, and no browser-based tool can recover it. The same goes
for EXIF/XMP fields and PDF document metadata: to check those you need the original file, not a copy of
its contents.