Supported hidden markers
None foundPaste text to scan it locally.
Free browser-local tool
Clean invisible text from AI chat output
AI assistants and chat interfaces can leave invisible content in copied text: zero-width characters, Unicode tag sequences, bidirectional controls, byte-order marks, word joiners, and unusual spaces. Anthropic has said it will ship invisible text watermarking to Claude, which is a statistical pattern across the token distribution rather than a visible or inserted character. This free tool makes supported hidden characters visible in a category report and can remove them before you copy the result, which also cleans common invisible markers that appear in AI and pasted text. It does not claim to detect, remove, or prove the absence of such a statistical token watermark. The paid AI Rewrite uses one credit to rephrase your text with clearly different wording and structure; rewriting is the only known practical way to interfere with a token-distribution signal, and no tool can certify that outcome. The cleaner runs locally in your browser: ordinary letters, punctuation, line breaks, emoji and meaningful joiners are preserved, and the text is never uploaded. AI Rewrite is a separate signed-in server task that reads your text once to produce the rewritten copy and charges credits from your balance.
Paste text to inspect hidden characters
Find supported zero-width characters, Unicode tags, abnormal spaces, bidirectional controls and invisible text markers.
AI rewrite · 1 credit
Server task · credits requiredRewrite with AI
Uses a paid provider to rephrase copied or AI text with clearly different wording. Sign in to use credits; the free cleaner above needs no account.
Rewriting changes your wording and is best-effort: it is not a detector, does not guarantee removal of any token-level watermark, and must not be used to pass authorship, fraud, or platform checks.
Supported removal types
Supported invisible text categories in AI output
The cleaner recognizes a deliberately narrow set of code points so its report is explainable. These include U+200B zero-width space, U+2060 word joiner, embedded U+FEFF byte-order marks, Unicode tag characters in the U+E0000–U+E007F range, supported bidi embedding, override and isolate controls, and unusual spacing characters that can be normalized to an ordinary space. Counts are shown before any change is applied.
Copied AI text can contain these markers from formatting, tracking, chat-context injection, or platform tooling. Removing them changes the code-point sequence and can make the text easier to compare, search, or reuse. The free tool does not rewrite meaning or change wording; the separate paid AI Rewrite step rephrases your text with different wording while preserving meaning.
- U+200B zero-width space, U+2060 word joiner, and embedded U+FEFF
- Unicode tag characters in the supported U+E0000–U+E007F range
- Supported bidi embedding, override, and isolate controls included by Text Clean
- No-break, thin, narrow, ideographic, and other supported unusual spaces normalized by the preset
- Paid AI Rewrite for wording-level rephrasing through a separate provider
How it works
Inspect AI text before copying the cleaned result
The text remains in browser memory and is never sent to a server; the tool needs no file upload, account, or network processing.
- 01
Paste or type text from Claude, another chatbot, or any source into the input. Nothing is removed while you are still reviewing it.
- 02
Inspect the category counts for supported hidden characters found in the input.
- 03
Review the fixed Text Clean preset, including its supported bidi and spacing rules, and keep the original where those code points may be meaningful.
- 04
Run the clean and compare the category totals, line breaks, and visible text before copying.
- 05
Keep the original when you need to verify meaning, direction, or byte-for-byte content in legal, multilingual, signed, or hashed material.
Metadata vs pixels
A hidden character is not image metadata or a pixel watermark
A hidden Unicode character is part of the text string. It can affect searching, comparison, cursor movement, line wrapping, or direction even though it has little or no visible width. Cleaning it changes the sequence of code points, not metadata attached to a separate file.
Image metadata lives in EXIF, XMP, IPTC, C2PA, or format-specific chunks, and an invisible image watermark is encoded into pixels. This text page handles neither. The watermark Anthropic has described for Claude is a statistical pattern across generated tokens; it is not a Unicode character inserted into the output and is not visible in the text. A character cleaner and a token-pattern fingerprint are different layers and must not be treated as the same thing.
What is preserved
What the safe text preset keeps
Visible letters, digits, punctuation, symbols, tabs, and line breaks remain in order. The cleaner keeps zero-width joiner, zero-width non-joiner, variation selectors, combining marks, and ordinary directional characters by default because automatic deletion could break scripts, emoji, names, or technical content.
Cleaning does not paraphrase, spell-check, translate, summarize, or alter letter case. Supported unusual spaces are normalized to ordinary spaces and reported separately. The output report shows category and removal totals so a user can notice an unexpected difference before copying.
Limitations
Language, direction, and watermark limits
Removing bidi controls can change the displayed order of mixed right-to-left and left-to-right text. Normalizing spaces can alter line wrapping or byte-for-byte signatures. Joiners and selectors may be invisible but semantically important. Review multilingual, source-code, domain-name, legal, and cryptographically signed material with the appropriate specialist tools before replacing the original.
Anthropic's announced textual watermark for Claude is a statistical fingerprint over the token distribution, not an inserted character. Character-level cleaning cannot detect or prove the removal of that fingerprint, and no tool can guarantee such an outcome. The paid AI Rewrite changes the wording itself, which is the only practical known intervention. This site does not claim to make any text undetectable or to defeat a detector.
- Not a detector for Claude, SynthID, or any proprietary watermarking system
- Character cleanup is not an AI-detection bypass; AI Rewrite changes wording but offers no detector guarantee
- Not a guarantee of authorship, originality, or platform acceptance
- Hidden markers in other copies or platform-side signals are outside this tool
Reproducible example
Test-fixture example: category counts and a stable visible result
A literal Unicode fixture makes every expected change reproducible without using private user text or claiming any connection to a proprietary detector.
Input
- The sequence “alpha<U+200B> beta<U+2060> gamma *name*” with escaped markers replaced by the actual code points
- One Unicode tag sequence, one narrow no-break space, one bidi isolate pair, and the emoji family sequence joined with U+200D
- Expected counts recorded for zero-width, tag, bidi, unusual-space, and protected-joiner categories
Change
- The Text Clean preset removes U+200B, U+2060, and the tag sequence and normalizes the supported unusual space.
- Bidi controls are removed only in the explicit bidi test; U+200D joiners remain in every default-preset run.
Verification
- The output equals the fixture's expected code-point sequence, not merely the same-looking rendered text.
- Category totals match the number of removed, normalized, and preserved characters.
- No claim is attached about any model's watermark, a detector outcome, or text provenance.
FAQ
Questions about this tool
Is Claude adding a text watermark?
Anthropic has said it intends to ship invisible text watermarking to Claude. The described technology is a statistical pattern across the token distribution, not a visible character or inserted unicode marker. This page treats Claude separately from supported hidden-character cleanup and makes no claim about that fingerprint.
Can this tool remove the Claude watermark?
The free cleaner removes supported hidden characters that appear in copied or AI text. It cannot detect or prove the removal of a statistical token-distribution watermark, because such a mark is not a character in the file. The paid AI Rewrite rephrases your text with different wording and structure — the only known practical intervention — but no tool can certify that a watermark is gone.
How do I use the paid AI Rewrite?
Sign in, paste at least 40 characters, choose Light, Medium, or High intensity and confirm ownership, then submit. One rewrite costs 1 credit from your balance; new accounts receive 3 welcome credits. A confirmed provider failure returns the credit automatically.
What hidden characters may appear in AI or pasted text?
Copied text can include zero-width spaces and joiners, Unicode tag sequences, byte-order marks, bidirectional controls, and unusual spaces. Text Clean makes these visible by category and removes its documented supported set on demand.
Does cleaning change the meaning of my text?
No. The preset removes its documented hidden-character set and normalizes supported unusual spaces. It preserves letter case, wording, order, line breaks, emoji sequences, and meaningful joiners.
Is the text sent to a server?
No. The free cleaner runs in the current browser and does not send the entered text to perform the operation. Clear the input when you finish on a shared device.
Related tools and guides
Continue with a verified path
Ownership notice
Use the text cleaner only on material you are permitted to edit. Do not use hidden-character removal to alter evidence, signatures, official records, another person's statement, or content for fraud, impersonation, or detector evasion.