Skip to main content
AI Watermark Removal

AI Text Forensics

Hidden Unicode AI Watermark: Proves Almost Nothing

Invisible characters turn up in AI chat output constantly, and almost every one of them is a formatting artifact that predates chatbots by years or decades. No major provider has documented a watermark built on zero-width spaces, joiners, byte-order marks, or non-breaking spaces, and Anthropic has now ruled the idea out for Claude in writing: "Nothing is added to the text and there are no hidden characters." OpenAI says the same of textGrain, the text watermark it announced on 2026-10-05 for opt-in API text and EU ChatGPT and Codex text: it inserts no hidden characters or invisible spaces. Homoglyphs are the one category that deserves a second look, and no provider has documented using them as a watermark either.

By ignacio-sjPublished Revised Sources verified Community discussion

Correction,

Two claims were wrong. The page described this site's cleaner as reporting counts for eight characters and said a correct cleaner deletes the zero-width family while replacing the two non-breaking spaces with an ordinary space. The cleaner covers more than 50 named characters plus several ranges in four tiers, and decides zero-width spaces, joiners and both non-breaking spaces by context, because deleting U+200B destroys word boundaries in Thai, Khmer, Lao and Myanmar and folding U+00A0 scrambles numbers in Hebrew or Arabic. The page also endorsed a forum claim that single-sample detection of a statistical watermark is effectively impossible; these schemes score one passage, and OpenAI reports about 80% detection at 200 tokens and 95% at 400. Passage length is the limit, not sample count. A blogger's claim that Claude Code once hid Unicode homoglyphs in date strings has also been removed, because no source for it could be linked.

All corrections

Key takeaways

  • Four characters explain nearly every find: zero-width space (U+200B), zero-width joiner and non-joiner (U+200D and U+200C), byte-order mark (U+FEFF), and non-breaking space (U+00A0). Each has a mundane, documented purpose unrelated to AI.
  • Zero-width joiners and non-joiners are load-bearing for Arabic and other complex scripts, and the joiner is what builds compound emoji. Byte-order marks are usually left behind by whatever software last saved or exported the text.
  • Homoglyphs are the genuine exception. A Cyrillic character standing in for a Latin lookalike has no everyday formatting purpose, and security researchers already treat character swapping as a real obfuscation technique.
  • Community threads reacting to Claude's 2026 watermark confirmation listed hidden Unicode as one guess among six. Anthropic settled it on August 14, 2026: the mechanism is keyed word selection, a SynthID-Text variant, and "Nothing is added to the text and there are no hidden characters." Separate marketing claims about hidden Unicode in Grok's text still have zero corroboration from xAI.
  • Finding one of these characters doesn't identify which model wrote the text, or that a model wrote it at all, unless a provider has published documentation of a scheme built on that exact character. None has.
  • Zero-width steganography is a real, decades-old technique: a zero-width joiner (U+200D) stands for one bit and a zero-width non-joiner (U+200C) for the other, building a payload with nothing rendering on screen. It decodes exactly from a few words and dies to one find-and-replace. Statistical watermarks are the opposite: they survive light editing but need a long passage to detect.
  • We measured it rather than assuming it: across 128 outputs and 26,093 words from four Claude models, including Fable 5.1, zero zero-width or bidirectional-format characters appeared, verified against a positive control proving the capture path preserves them. The original run covered 96 outputs from three model tiers, and version 1.1.0 added Fable 5.1. The only invisible characters found were 48 ideographic spaces inside Japanese-language output, which is ordinary Japanese typography.

Figure 1

Invisible character diagnosis: where each one actually comes from

Every character on this list predates chatbots by decades. Finding one tells you something about how the text was copied, not about which model produced it.

CharacterCommon sourceVisible effectSafe to remove?Evidence of a watermark?
Zero-width spaceU+200BCopy-paste from web pages, CMS line-break hintsNoneYesNo
Zero-width joinerU+200DEmoji sequences, Indic and Arabic scriptsNone, but load-bearing in some scriptsCarefulNo
Byte-order markU+FEFFFile encoding headers, exports from ExcelNoneYesNo
Non-breaking spaceU+00A0Word processors, HTML entitiesLooks like a spaceUsuallyNo
Narrow non-breaking spaceU+202FFrench typography, unit spacingThin spaceUsuallyNo
Soft hyphenU+00ADJustified text, PDF extractionNone until line-wrappedYesNo
Word joinerU+2060Typesetting, deliberate no-break pointsNoneYesNo
HomoglyphsCyrillic а, Greek οMixed-script paste, phishing, deliberate obfuscationIdentical to Latin lettersCarefulNo

The column that matters is the last one. Removing these characters cleans up text that will be diffed, searched, or parsed. It does not touch a statistical watermark, because a statistical watermark is not made of characters. It lives in which words the model chose.

Method Sources are the ordinary provenance for each character in normal document handling. The watermark column reports whether any provider has documented that character as part of a marking scheme. None has.

Checked 2026-08-13

The 128-output scan

Which invisible characters turn up most often in AI text?

Confirmed

Here's what is actually sitting in your clipboard, and why each character was there long before any chatbot existed.

Four characters account for the overwhelming majority of invisible-character finds in AI output.

  • Zero-width space (U+200B): no glyph, no width, used to create line-break opportunities inside long unbroken strings like URLs and hashtags.
  • Zero-width non-joiner (U+200C) and zero-width joiner (U+200D): control whether adjacent letterforms connect in Arabic and other complex scripts.
  • Byte-order mark (U+FEFF): written at the start of a file to signal byte order or confirm UTF-8 encoding.
  • Non-breaking space (U+00A0): keeps two words on the same line, like a name and a title.

Every one of those predates the modern chatbot, in some cases by decades. And each has an ordinary explanation for how it ended up in your text.

The zero-width space does have a less charitable use. Some content generators and spammers sprinkle it to pad word counts or slip past duplicate-content filters, which is a spam signal, not a provider signature.

The joiner is the one people underestimate. It's what glues emoji sequences together, so the family emoji and the profession sequences are literally built out of U+200D. Skin tones are different: they use separate modifier characters, U+1F3FB to U+1F3FF, attached directly with no joiner.

A byte-order mark is an encoding artifact left by whatever editor, export tool, or paste pipeline last handled the text. It tells you something about the software that saved the file and nothing about who wrote the words inside it.

The non-breaking space is the most common false alarm of the lot. Enormous numbers of web pages use it in their markup in place of a plain space, so any text copied from a browser is likely to carry a few.

We counted these characters directly in Claude's output: the invisible-character census scanned 128 outputs and 26,093 words from four Claude models, including Fable 5.1, and found no zero-width or bidirectional-format characters. The original run covered 96 outputs from three model tiers, and version 1.1.0 added Fable 5.1.

How does zero-width space steganography hide a message, and how is it different from a statistical watermark?

Confirmed

Hiding a message in invisible characters is a real technique. It works in the opposite way from the watermarks providers actually ship.

A zero-width message is built one bit at a time out of characters that render nothing on screen. The zero-width joiner (U+200D) and zero-width non-joiner (U+200C) are the usual choices: one stands for a 1 and the other for a 0.

Slot them between ordinary letters and a program that knows the scheme reads the sequence straight back out. This is general-purpose text steganography and it long predates modern AI.

The zero-width space (U+200B) is the character people usually suspect first. It has no glyph and no width, and it only marks a place where a line may break.

A statistical watermark does not add a character. SynthID Text runs as a logits processor during generation: after ordinary Top-K and Top-P sampling narrows the candidate next tokens, it reshapes which of them the model is likely to pick, so the signal is spread across dozens of word choices.

Anthropic describes Claude's watermark as "a version of the SynthID-Text approach published by Google DeepMind," with a secret key and the preceding words determining which eligible next word gets picked.

How easily does a zero-width space get lost? The Lab's transformation-survival study pushed 18 invisible characters through 12 ordinary transformations: U+200B survived 9 of the 12, including Unicode normalization, JSON round-trips, URL encoding, Base64 and line trimming, and was removed only by stripping Unicode format characters, by Latin-1 storage, and by stripping all non-ASCII text.

Do Unicode homoglyphs prove that AI wrote the text?

Community discussion

A Cyrillic letter posing as a Latin one has no innocent formatting excuse, but no provider has documented using one as a watermark.

No, though homoglyphs are the one category worth a second look. A homoglyph swaps a letter for a visually identical character from another script, like a Cyrillic character standing in for a Latin a. Your eye can't catch it. A Unicode-aware tool can.

Unlike a byte-order mark or a non-breaking space, there's no everyday formatting reason to do this. Security researchers already treat homoglyph substitution as a real technique, used in phishing domains and text obfuscation.

No provider has documented a watermark built on homoglyphs, and Anthropic's August 14, 2026 explainer rules out added characters of any kind for Claude: "Nothing is added to the text and there are no hidden characters." A homoglyph is a substituted character, so it falls under the same statement.

So a lookalike character in your text is worth investigating as a sign of copy-paste from another script or of deliberate obfuscation, but it does not identify a model, and it does not show that a model wrote the text.

Does Claude's watermark use hidden Unicode characters?

Community discussion

Six guesses about Claude's mechanism circulated; on August 14, 2026 Anthropic named the real one, and hidden Unicode was not it.

No, not according to Anthropic, which ruled hidden characters out in writing on August 14, 2026. When Anthropic confirmed a Claude text watermark in 2026, the same follow-up question filled every comment section: how does that even work for text? The theories piled up fast.

  • hidden Unicode or invisible characters
  • statistical word-choice patterns
  • overrepresented n-grams
  • first-letter or sentence-position patterns
  • token-probability nudges
  • something SynthID-like, or a hybrid of several signals at once

For twelve days Anthropic's own documentation named none of them, and hidden Unicode, the only guess testable at home, got checked first and reported loudest. On August 14, 2026 Anthropic ended the guessing: the watermark is a version of Google DeepMind's SynthID-Text approach, keyed word selection during generation, and "Nothing is added to the text and there are no hidden characters." That settles this page's question from the provider's side: there is no hidden character to find.

The same pattern shows up elsewhere. Marketing pages for the removal tools gptcleanup.com and scriptbyai.com have claimed hidden Unicode in Grok's text output, a claim with zero corroboration from xAI.

Compare that with getgpt.app, whose scanner checks for 34 or more specific invisible characters and states plainly that it does not detect or remove cryptographic watermarks like SynthID. Same underlying capability, honest label.

What should you do with an invisible character you found?

Confirmed

Strip them, for the ordinary reasons: broken search, corrupted pastes. Just don't mistake that cleanup for defeating a watermark with no character to delete.

Stripping invisible characters is worth doing on its own merits. They break search-in-page, corrupt pastes into code and spreadsheets, and occasionally trip duplicate-content checks.

This site's cleaner runs entirely in your browser and sorts more than 50 named invisible characters, plus the variation-selector, tag and private-use ranges, into four tiers. Byte-order marks, soft hyphens and invisible math operators are removed. Zero-width spaces, joiners and non-joiners and the two non-breaking spaces are decided by the characters either side. Exotic widths such as the em space are normalised to an ordinary space. Directional controls, the word joiner (U+2060) and the ideographic space are reported and left alone. Nothing is uploaded anywhere.

What it won't do is touch a statistical watermark, because there's no character to delete. Cleaning invisible Unicode and defeating a watermark are separate jobs, and only one of them is available to you in a text box.

FAQ

If I find a zero-width character in text I copied from an AI chatbot, does that prove which model wrote it?

No. Zero-width spaces, joiners, byte-order marks, and non-breaking spaces all have ordinary technical explanations with no connection to any specific AI provider. Unless that provider has published documentation of a watermark built on that exact character, its presence identifies nothing, and doesn't even confirm the text is AI-generated.

Are lookalike Unicode characters as harmless as zero-width spaces?

No, and it's worth separating the two clearly. Homoglyphs have no everyday formatting purpose, which is why security researchers treat character swapping as a real obfuscation technique. But no provider has documented a watermark built on them, and Anthropic says nothing is added to Claude's text.

Does finding a hidden character at least confirm the text passed through some AI tool?

Not reliably. Word processors, CMS pipelines, and ordinary web copy-paste all insert these characters with no AI involved. A stray zero-width space or byte-order mark is a weak signal of almost nothing, let alone proof of which system produced the text.

Has any provider ever documented a Unicode-based watermark?

Not one, and Anthropic has now documented the opposite: its explainer of August 14, 2026 says "Nothing is added to the text and there are no hidden characters." Google documents SynthID Text as a change to token sampling during generation, and Anthropic describes its watermark as a version of the same approach. OpenAI's textGrain, announced on 2026-10-05, also works through word choice, and OpenAI says it inserts no hidden characters or invisible spaces. None of the three inserts a character you could find and delete.

How does zero-width steganography actually hide a message?

By encoding bits as invisible characters. A common version inserts a zero-width joiner (U+200D) after certain letters to mean one bit and a zero-width non-joiner (U+200C) to mean the other, building a payload character by character with no glyph rendering on screen. It's a decades-old idea that predates modern AI entirely.

Does removing zero-width characters count as removing an AI watermark?

No. It's useful text hygiene, fixing broken search indexing and copy-paste artifacts, but no major provider has documented zero-width characters as its watermarking mechanism. Google's SynthID Text and Anthropic's Claude watermark both work by shaping which words a model picks during generation, not by inserting an invisible character afterward.

Next steps

  • See the measurement behind this: 128 Claude outputs from four models, including Fable 5.1, scanned character by character, with the data and the script published. The invisible-character census
  • Run a suspicious passage through the free in-browser cleaner and see exactly which characters are in it, with counts per character. Scan your text
  • Read how statistical watermarking bends token probabilities during generation, which is the mechanism zero-width theories keep getting mistaken for. How statistical watermarking works
  • Check what Anthropic has and hasn't said about Claude's real, confirmed text watermark. The Claude watermark, sourced
  • If it was punctuation rather than an invisible character that made you suspicious, start with the em dash myth. Do em dashes mean AI?

Sources and citation status