Claude watermark detector, no key

Two quantities, computed locally, reported separately. The first is an occurrence count of 14 specific invisible codepoints present in the submitted text. The second is a count of non-overlapping spans matching 17 writing patterns. Each finding is rendered in place with the reason attached to it, so the report can be read item by item rather than as a total.

The pass is local and deterministic. Nothing is transmitted, no allowance is metered, no character cap applies, and scanning alters nothing. A cleaning action is available in the same widget, but inspection and mutation are separate operations here on purpose.

Anthropic's help documentation describes an imperceptible watermark placed directly into the text, which travels with the text when it is copied and pasted, and states that watermark detection is currently in private preview rather than publicly available. Access is restricted to eligible organisations under EU law, named as regulators, law enforcement, media, fact checkers, independent researchers, educational organisations and EU civil society groups, plus enterprises with compliance obligations.

The mark is read with a key that encodes it. No third party holds that key, so no third-party tool performs that check, and this page does not claim to. Anthropic's separate free checker, open to everyone, operates on files: it reports whether a file carries a Claude-issued Content Credential under the C2PA standard. That is a provenance question about a file, not a question a text field can resolve.

Inspect the artefact in its received state. Reformatting or retyping before the scan destroys precisely the evidence the scan exists to surface.

Step 1

Paste the text unmodified. No length cap applies. A sample loader is available to populate the report with known dirty input.

Step 2

Scan. Read the two counters as separate results and open each marked finding for its individual reason.

Step 3

Decide per finding. Cleaning is a distinct action that removes characters only, and it is not required in order to have inspected.

The character count is a fact about the string. Those codepoints are present at those offsets, and the count is reproducible by anyone with the same input. It supports no inference about who produced the text, because its cause is the copy path: rendered interfaces, rich-text clipboards, word processors inserting no-break spaces around numerals.

The phrase count is a different object entirely. It measures how far the prose falls into recognisable patterns, and the matches are de-duplicated before counting: candidates are ordered by position, longest first, and any match starting inside a previous one is dropped. So the figure is spans of affected text rather than raw regex hits.

Neither is evidence of authorship, and they should never be summed. One tells you what is in the bytes. The other tells you how the sentences read.

The list is fixed in this codebase. It is not user-extensible and it is not inflated to make the page look thorough.

Zero-width group, four

Zero-width space, non-joiner, joiner, word joiner. Each occupies a slot while rendering nothing, so any exact comparison sees a string other than the displayed one.

Byte-order mark

Legitimate at file head, disruptive mid-document, where it produces parser and validator failures that look like data corruption.

Soft hyphen

A conditional hyphen, inert until wrapping, at which point a hyphen materialises inside a word that never contained one.

Space look-alikes, three

No-break space, narrow no-break space, hair space. Visually identical to U+0020 and distinct to every comparison operation.

Bidirectional controls, five

Two embeddings, a pop, and both overrides. An override decouples display order from storage order, which has been used to make filenames and links read as something they are not.

The exclusion

A zero-width joiner flanked by pictographic characters is composing an emoji sequence and is not reported. Reporting it would produce a false finding on every composed glyph.

An audit of the tools ranking for this term makes the sharpest available point about them: stripping zero-width characters produces a clean, reassuring result on text that is fully watermarked, because the operation was never looking at the watermark. The same logic applies to a clean scan on this page, and stating it is the reason to trust the rest of the report.

Anthropic frames the symmetric limitation on their own side. A detected mark is described as a signal that content may have been processed by Claude and as not fully conclusive, and the absence of a mark does not mean the content was not AI-generated or processed. A negative result is evidence about the test, not about the text.

Not authorship, not model attribution, not probability of generation. No score, percentage or confidence figure is produced anywhere on this route, because nothing in the path computes one.

It does not estimate how much watermark a passage could carry, which the same audit identifies as what competing checkers are actually computing under the detection label. It does not inspect punctuation, so smart quotes, en dashes and em dashes are neither reported nor counted, and it does not inspect markup. There is no file path, no C2PA reading, no batch mode and no API, and no prediction is offered about any third-party detector's output.

The audit of six ranking tools found systematic misdescription: pages advertising detection of a mark that no public tool can read, one describing it as invisible characters, another listing HTML fingerprints and AI-style punctuation among its signals. The more careful tool pages in the set disclaim it in their own FAQs while keeping detection in the headline.

There is a reasonable version of this page, and it is the narrow one. State the two things that can be measured from pasted text, attach a reason to every finding, and name the question that stays out of reach until Anthropic's detection becomes generally available.