LLM tools

Private browser tool

Remove hidden marks from AI text

Paste text from an LLM or document. Supported, structurally detectable C2PA text forms are removed as it lands. Suspicious mixed-spacing sequences and other hidden formatting marks are opt-in—no clean button, upload, or account. This tool does not detect and is not designed or verified to remove Claude's model-level watermark.

The important distinction

What this tool can—and cannot—remove

It does remove

C2PA-identified Unicode wrappers, conservative edge/front-matter manifest references, and recognized C2PA associations in an explicit HTML head. When you opt in, it can also remove selected hidden formatting characters and neutralize runs that mix zero-width spaces, hair spaces, and narrow no-break spaces while leaving those typographic spaces alone when they appear independently.

It is not designed to remove

Claude's keyed, SynthID-Text-based statistical watermark. Anthropic says it creates a keyed pattern through the words Claude chooses; nothing is added and there are no hidden characters. Its public detection API is forthcoming. A result with no matching removable form is not proof that text was written by a human or that every possible watermark is gone. Removing an embedded association also cannot prove that a durable Content Credential is undiscoverable through another binding.

Research, updated 15 August 2026

Can this remove Claude AI watermarks?

This cleaner does not detect and is not designed or verified to remove Claude's model-level watermark. Anthropic says Claude models launched on or after 2 August 2026 support marking at launch, using a version of Google DeepMind's SynthID-Text. Anthropic is working to add marking to older models over the coming months. The method uses a secret key and preceding words to create a statistical pattern through low-stakes word choices. Anthropic explicitly says nothing is added to the text, there are no hidden characters, and the watermark carries no identifier for a user, organization, or chat. This cleaner's C2PA and Unicode cleanup is not designed or verified to remove that signal.

Anthropic says detection estimates the likelihood that Claude was involved; it is weaker for short, factual, lightly proofread, or code-heavy passages. Light editing may leave the signal, while a complete rewrite that replaces every word can remove it. Rewriting changes the visible prose and is outside this cleaner's meaning-preserving scope. Detection requires the watermark key; Anthropic says a detection API is coming soon, but it is not yet publicly available.

Anthropic separately describes signed C2PA provenance metadata for supported generated files such as PNG, JPG, and SVG. That file-attached metadata is different from its embedded text watermark. When you copy ordinary rendered text from a file into a text box, the file container and its binary metadata do not come along for the ride.

What C2PA text metadata looks like

C2PA 2.4 defines an under-review method for provenance in plain text. It starts with a zero-width no-break space, then encodes a wrapper whose first bytes spell C2PATXT\0 using Unicode variation selectors. The standard also defines visible manifest blocks for structured text and dedicated script or link associations for raw HTML. This cleaner conservatively checks beginning/end comment associations, front matter, explicit HTML heads, plus the wrapper header and outer JUMBF C2PA identity before reporting a C2PA-identified Unicode wrapper. It does not fully validate the manifest or its signature, or delete every invisible character indiscriminately.

That restraint matters. Zero-width joiners and non-joiners can be meaningful in Persian, Arabic, Indic scripts, and emoji. Variation selectors can choose emoji presentation or a specific CJK glyph. Those characters are preserved unless they are part of a recognized C2PA wrapper. Recognized emoji flag tag sequences are preserved too.

The opt-in pass uses one narrow heuristic: a run that contains all three of a zero-width space, hair space, and narrow no-break space. It replaces that mixed run with at most one ordinary word boundary. It does not globally erase hair or narrow no-break spaces, which are legitimate typography in numbers, units, and punctuation. The match does not identify a provider or prove that the sequence encodes a watermark.

How to clean pasted LLM text online

  1. Tap the text area and paste the content. Recognized C2PA cleaning happens synchronously as part of the paste.
  2. Only if needed, enable hidden formatting cleanup to neutralize suspicious mixed-spacing sequences and review its warning before copying.
  3. Review the category-level removal summary. “No matching hidden formatting found” is intentionally not a provenance claim.
  4. Tap Copy cleaned text. The copy call starts directly from your tap for mobile Safari compatibility.

Cleaning is local to this browser tab. This tool makes no network request with your text, stores no text on a server, and records no analytics event containing the pasted content. Spellcheck and autocorrect are disabled on the field; your browser, keyboard, or extensions may still have their own privacy settings. Closing or refreshing the page clears the working text.

Why the claim is deliberately narrow

“AI watermark” can mean cryptographic file provenance, invisible Unicode, or a statistical pattern in a model's token choices. Those mechanisms are not interchangeable. Anthropic says light editing may leave Claude's signal while a complete rewrite that replaces every word can remove it; if Claude produces a translation, that new output receives its own watermark. Rewriting changes the writing and is outside this narrow, visible-prose-preserving pass. Legal and disclosure duties vary by jurisdiction and use; this utility is a text-sanitization tool, not legal advice or a provenance certificate.

Does the cleaner rewrite the visible text?

The default pass does not paraphrase sentences, normalize spelling, replace punctuation, or change letterforms. Ordinary newlines, tabs, emoji, accents, and non-Latin scripts stay intact. The optional hidden-formatting pass can change hyphenation, line breaking, hidden tag data, or an identified mixed-spacing run, so it starts off and carries an explicit warning. Standalone typographic spaces remain unchanged. Supported C2PA-form cleanup remains active either way.

Primary sources