Back to list
AI5 min readUpdated 11/2/2026

What Are AI Watermarks? Why Should You Remove Them?

Discover the invisible watermarks hidden in AI-generated text from ChatGPT, Claude, and others, and learn why removal might be necessary.

By webgyver

What Are AI Text Watermarks?

Invisible characters sometimes called "AI text watermarks" can show up in AI-generated text. They typically exist in the form of zero-width characters, Unicode variation selectors, and bidirectional text control characters.

These watermarks are completely invisible to the naked eye but actually exist within the text data and can cause various problems. Such characters have been reported in some AI-generated text, but major AI services have not been confirmed to offer them as an official feature. The same characters can also be introduced by copy-and-paste or editor conversions.

What Types of Hidden Characters Exist?

Zero-width characters (ZWSP, ZWNJ, ZWJ) are the most common form. These characters don't display on screen but are included in text length, interfering with character count calculations.

Unicode variation selectors are used to change how characters are displayed but can be misused for watermarking purposes. Bidirectional text control characters (LRM, RLM, etc.) work similarly.

Combining diacritical marks are used to add accents or symbols above base characters but can be used as invisible marks.

Why Should You Remove AI Watermarks?

First, they cause problems when accurate character counting is needed. When posting to platforms with character limits, more characters may be counted than what's actually visible.

Second, they can cause unexpected errors in code or data processing. Issues can arise in string comparison, searching, and sorting operations in programming.

Third, from a privacy perspective, you may not want traceable identifiers included in content you've created. Maintaining clean text is part of digital privacy.

How to Use TheWebGyver's AI Watermark Remover

The usage is very simple. Just copy and paste the text received from AI into the input field. The tool automatically detects and displays hidden characters.

Detected watermark characters are categorized and displayed, and all hidden characters can be removed with a single click. The processed clean text can be copied immediately.

All processing occurs within the browser, so text is never sent to external servers. You can safely process personal and sensitive content.

Myth vs. Reality: Invisible Characters and AI Watermarks Are Not the Same Thing

Now and then people report finding zero-width characters after copying an AI answer, and three misconceptions tend to follow. The first is that any invisible character you find must be a watermark planted by an AI. A zero-width space (U+200B) or a byte order mark (U+FEFF) can enter text through web page copying, document editors, messengers and file format conversion, so finding one does not establish where the text came from.

The second misconception is that AI services therefore insert such characters. No major AI service has been officially confirmed to insert zero-width characters as a watermark. The research literature does list the approach as a category, though. The SynthID-Text paper from Google DeepMind, published in Nature in October 2024, divides text watermarking into generative watermarking, edit-based watermarking applied to already generated text, and data-driven watermarking that alters training data. As examples of the edit-based kind it names synonym substitution and inserting special Unicode characters. Being listed as a category is very different from being used by a real service.

The third misconception is that a watermark must be a hidden character. The SynthID-Text scheme described in that paper inserts no characters at all. It modifies only the sampling procedure by which the model picks the next token, and the paper says detection is computationally efficient and does not use the underlying model. The 2023 ICML work by Kirchenbauer and colleagues proposes selecting a randomized set of green tokens before each word is generated and softly promoting them during sampling, with detection that needs no access to the language model API or parameters. A statistical watermark like this contains no special characters anywhere. It exists only as a pattern in word choice.

Characters to Remove and Characters to Keep: A Checklist by Situation

Removing every hidden character is not always the right move. In the Unicode character database (UnicodeData.txt), U+200B (zero width space), U+200C (ZWNJ), U+200D (ZWJ), U+200E and U+200F (directional marks), U+2060 (word joiner) and U+FEFF are all classified as format characters, category Cf. None of them is drawn on screen, but they do different jobs.

The first situation is plain Korean or English prose. U+200B, U+2060 or U+FEFF sitting between letters can usually be cleaned up without harm, and they are more likely to cause trouble in length-limited input fields or string comparisons.

The second situation involves emoji. The Unicode emoji standard (UTS #51) uses ZWJ between characters as a request to show a single glyph if one is available, and such emoji sequences are treated as a single grapheme cluster. The variation selector U+FE0F, which requests emoji presentation, is described in the same standard. You can see the effect by running it. In Node.js v26.12.1, a family emoji made of man, woman and child joined by two ZWJs had a UTF-16 length of 8 and one grapheme cluster according to Intl.Segmenter. After removing only the ZWJs with a regular expression, it split into three grapheme clusters, three separate emoji. That is why deleting hidden characters blindly can change how emoji look.

The third situation is text that mixes in right-to-left scripts such as Arabic, Persian or Hebrew. The Unicode FAQ mentions that ZWNJ and ZWJ affect cursive joining and ligation, and describes the right-to-left mark (RLM) as an invisible character used in bidirectional text. In such languages these characters may be meaningful, so check how the result displays before removing them in bulk.

Step by Step: Checking Suspicious Text Yourself

Step one is to compare the number of characters you can see with the number a program counts. As an example, a string made of the five Korean syllables of annyeonghaseyo with two U+200B characters between them had a length of 7 when checked in Node.js v26.12.1. It looks like five characters but counts as seven, which explains how text can trip a character limit. This is a small hand-made example, not statistics gathered from real AI answers.

Step two is to know the limits of common cleanup methods. In the same environment, JavaScript's trim() removed a leading and trailing NBSP (U+00A0) and U+FEFF, but a string wrapped in U+200B stayed at length 4. In UnicodeData.txt, U+00A0 is a space separator (Zs) while U+200B is Cf. That is why running a whitespace-trimming function and assuming the text is clean is risky.

Step three is to see which characters are present and where, grouped by kind. You can turn on a code editor's option to show invisible characters, or print the string as code points (U+XXXX). TheWebGyver's AI watermark remover also lists detected characters by category, so it can serve as a way to check what was actually in the text.

Step four is to compare again after removal. In the example above, stripping the range U+200B to U+200D, U+2060 and U+FEFF with a regular expression brought the length back to 5. After cleaning, check not only the count but also that emoji and right-to-left sentences still display correctly.

What Removing Characters Cannot Fix, and Using It Responsibly

A statistical watermark such as SynthID-Text does not change when you delete characters. The Nature paper says the scheme modifies only the sampling procedure, so the mark lives in a statistical bias in word choice rather than in special characters. It is therefore reasonable to infer that deleting every hidden character would not affect such a mark. That is an inference from the paper's description, not a check of any particular service's implementation. Conversely, finding no hidden characters is not proof that AI did not write the text. The same paper notes that no text detection method is foolproof.

Cleaning hidden characters is best understood as data hygiene: matching a character count, avoiding string comparison errors, or tidying pasted text. It is not a way to conceal that AI was used, and it makes no promise of doing so. Schools, employers and publishers have different disclosure rules for AI assistance, so follow the ones that apply to you and, where appropriate, say that AI helped.

When you handle sensitive writing, also check where the processing happens. A tool that works entirely inside the browser does not send your text to a server, but whichever tool you use, it is worth reading its notice before pasting.

References

Try It Yourself

Experience what you learned in this article with AI Text Watermark Remover.

AI Text Watermark Remover