AI Watermark Center

Text tool

AI Text Watermark Scanner

Check pasted text for hidden Unicode, invisible characters, zero-width marks, directional controls, special whitespace, and normalization differences. Use this checker when you need to inspect a string — for what those characters are, see the hidden Unicode guide.

This page owns transactional “hidden character checker / detector” intent. Anthropic states that Claude’s official text watermark is not based on hidden characters. This scanner checks Unicode artifacts only — it does not detect Claude’s statistical watermark or any SynthID text mark.

Processed locally in your browser. Your pasted text is not uploaded for analysis.

0 characters

Example text containing hidden Unicode characters. The sample is not from a named AI provider.

Paste text to inspect.

No inspection yet. Paste text and choose Analyze to list supported Unicode findings. A clean report is not proof the text is human.

What does the scanner inspect?

The scanner walks Unicode code points in the pasted string. It reports zero-width characters, unusual spaces, bidirectional controls, format characters, unexpected control codes, variation selectors, private-use code points, replacement characters, and whether the text changes under NFC, NFD, NFKC, or NFKD normalization.

Ordinary letters, combining marks used in real writing systems, tab, newline, and the normal space U+0020 are not treated as findings.

What are hidden Unicode characters?

Some code points have no ink, or look like an ordinary space while behaving differently. Copy and paste from word processors, websites, and messaging apps can carry them along. They are not a synonym for an AI watermark. This tool page is the checker for listing what is present in a paste; the related guide explains what those characters are and why they appear.

Can hidden characters prove text was AI-generated?

No. Hidden characters can originate from word processors, websites, messaging apps, formatting systems, copy and paste, legitimate language behaviour, or characters someone inserted on purpose. A clean report also does not prove that a person wrote the text.

Does this detect Claude’s watermark?

No. Hidden Unicode inspection is not Claude watermark detection. Anthropic states that Claude’s official text watermark is not based on hidden characters. This scanner checks Unicode artifacts only: hidden Unicode, zero-width characters, and formatting anomalies. It cannot inspect Claude’s statistical watermark, which Anthropic describes as depending on a provider key.

Current Claude detection options and the Claude watermark research hub explain the boundary in more detail.

Can the tool remove a watermark?

Supported Unicode cleanup is deterministic: the same input and policy produce the same output, and the after panel is a fresh scan. Conservative Cleanup only removes a leading U+FEFF byte-order mark candidate; it does not strip zero-width spaces, word joiners, soft hyphens, or emoji tags. That is not a claim that a provider-specific statistical watermark was removed. Absence of hidden characters does not prove absence of AI provenance.

Privacy

Inspection runs in your browser. Pasted text is not uploaded to AI Watermark Center for analysis. Future analytics, if enabled, will not include pasted text, excerpts, or finding context.

Methodology

Results follow the same rule as the rest of the site: we only claim what we can verify. Read the research methodology for the evidence hierarchy that later provider pages will use.

Related guides