Claude character detector

GPTCleanup Pro's Claude invisible character detector reads text copied from Claude and reports the zero-width and non-printing Unicode characters sitting inside it, each one named and located, rather than a vague warning about hidden watermarks. It's built for the question people type into search: why does Claude text have hidden characters. The answer here is a list of findings, not an accusation.

It's a server tool. Text is sent to our AI server to be scanned, and the result is returned to the page. You get 5 free checks in any rolling 24 hours, with no account required to run them.

This page treats a finding as evidence to be read carefully, not proof of anything. Hidden characters are common in text copied out of Claude's chat interface for reasons that mostly have nothing to do with a deliberate watermark, and the sections below explain why plainly.

One round trip to the server, then a specific list of findings. Nothing to configure beforehand and no particular format required.

Step 1

Copy the sample using Claude's own copy button, since that's the version most likely to carry hidden characters.

Step 2

Run the check. The text is sent once to our AI server, scanned for hidden Unicode, and nothing further is done with it after the result returns.

Step 3

Review each flagged character and its location. The result only reports what's there; it changes nothing in the text itself.

Claude's chat interface renders its replies as formatted, often markdown-heavy, rich text, and converting that to plain text on copy is where most of these characters enter the byte stream. Each one is an ordinary Unicode codepoint with a legitimate technical purpose:

U+200B zero-width space

Occupies a character slot and renders nothing, the most common invisible character in text copied from Claude.

U+200C / U+200D joiners

Zero-width non-joiner and joiner, occasionally a leftover from how Claude's rich text renders certain characters and emoji.

U+2060 word joiner

An invisible joiner preventing a line break at that point, often sitting near a dash or a formatted phrase.

U+FEFF byte-order mark

Meant to sit at the start of a file, occasionally left mid-response by an export or copy path.

U+00AD soft hyphen

Invisible until a line wraps there, then a hyphen appears in a word that never had one.

U+00A0 / U+202F non-breaking spaces

Visually identical to a normal space, distinct in code, slugs and search indexing.

Bidi controls (U+202A to U+202E)

Directional embedding and override characters, rarer in Claude text but worth checking for.

This is the point this page states plainly rather than leaving ambiguous. Despite widespread claims online that Claude embeds secret watermarks in its output, that isn't the usual explanation. These characters are overwhelmingly incidental artefacts of how the chat interface's rich-text editor tracks formatting and cursor position, and of the conversion that happens when that rich text is turned into plain text on copy.

The same kinds of characters appear when text is copied from a Word document, a Google Doc, a PDF or a translated page, none of which involve Claude at all. A finding here indicates something about that copy path. It does not, by itself, indicate the underlying words came from Claude rather than any other source, or that Anthropic marked the text on purpose.

No, and this tool does not claim it does. A hidden character is a formatting artefact of the copy path, not a fingerprint tied to Claude specifically. The same characters appear in human-typed documents exported from Word, in translated text and in PDF copy-paste, with Claude nowhere in the process.

It's worth distinguishing this from a separate, genuine area of research: statistical, token-level watermarking, in which a model's word-choice probabilities are tuned across a passage in a pattern detectable by the right statistical test, not by inserted characters. That technique operates on probability distributions, and no character-level scan, including this one, can observe it. A claim that Claude hides a watermark findable by scanning for zero-width characters describes something other than how real watermarking research actually works.

Several categories of ordinary, entirely human workflow produce the same characters as a Claude copy-paste, which is why a flag here isn't evidence of anything about who wrote the underlying words.

Word processor exports

Word, Google Docs and comparable tools insert non-breaking spaces and soft hyphens as part of standard formatting.

Translation tools

Machine and professional translation software frequently leaves directional or joining characters behind.

PDF extraction

Pulling text from a PDF commonly introduces stray characters the source document never displayed.

Other chat interfaces

ChatGPT, Gemini and other assistants render rich text as well, and can leave similar characters behind on copy.

CMS and rich-text editors

Toggling between rich-text and plain-text views is a routine source of leftover formatting characters.

Independent of origin, these characters cause concrete problems once text moves somewhere that parses it strictly. A zero-width space inside a heading splits a keyword, so a page's slug and its visible title stop matching. A byte-order mark stranded inside a JSON file breaks the parser, and non-breaking spaces confuse diffing tools and find-and-replace.

Pasting directly from Claude into a CMS, an email or a code editor is exactly where these problems surface. Checking first turns a category of confusing formatting bugs into a non-event.

Editors, developers and publishers pasting Claude output into a CMS, an email or a codebase use this to catch formatting problems before they surface downstream. It's server-side by necessity: your text has to travel to our AI server to be scanned, then the result travels back, which is why the free allowance is metered.

You get 5 free checks in any rolling 24-hour window, tracked without an account. That covers ordinary use, one response, one document, one flagged section at a time, and for a longer batch, spacing checks across the day works better than exhausting the allowance in one sitting.