Invisible Text Detector & Remover

Our Free Tool Identifies and Removes Invisible Unicode Text Characters:

This tool lets you easily visualize all hidden Unicode characters in your text and then remove them either strategically (by specific code) or all at once. 

How to use the tool:

1. Paste in your text
2. See all hidden Unicode characters in your text and information about what they do
3. Decide what you want to clean up
   
   1. Fix all
   2. Remove all invisible characters
   3. Strategically fix specific ones (like remove all Em-Dash and En-Dash) 
4. Copy your cleaned text

Note that if the text was created by an LLM (ChatGPT, Claude, Gemini etc) it will still be identified as AI-generated by Originality.ai’s AI Detector. 

Key Takeaways:

• One Click Cleanup - Our free Invisible Text Detector and Remover makes it easy to identify and remove any unicode hidden text characters
• Privacy Assured - All processing happens locally in your browser; your text never leaves your device.
• Not a Secret Watermark - LLMs like ChatGPT do inject hidden characters, but it is not for watermarking or nefarious reasons
• Overuse of Certain Characters - ChatGPT does use several Unicode “hidden” characters a lot (such as the Em Dash)
• Real Risks Exist - Invisible text characters can cause both security and formatting challenges 
• Does Not Bypass Detectors - AI Detection remains similarly effective whether the hidden characters are removed or not

All LLMs Do It - All popular LLMs inject hidden characters

What are Unicode Hidden Text Characters?

Hidden Unicode characters, whether subtly visible (em-dashes, smart quotes, no-break spaces) or completely invisible (zero-width spaces, joiners, direction marks) are special code points that don’t behave like plain ASCII. They act as a “digital ink” that reshapes how software wraps lines, splits words, parses data, or matches text, even when you don’t realize they’re there.

Does ChatGPT Watermark Text with Hidden Characters?

No - It has been incorrectly reported that the injection of hidden characters by ChatGPT is their attempt at “watermarking” their outputs. However,  there are 2 reasons that this is very unlikely..

1. It is incredibly easy to circumvent a watermarking strategy by just removing the characters. The ease at which users can do this would render a watermarking strategy pointless. 
2. OpenAI has stated it is just “a quirk of large-scale reinforcement learning.”

Does ChatGPT Inject Hidden Characters? 

Yes - LLMs like ChatGPT do inject hidden characters. Many of these characters are harmless formatting characters (like the popular em-dash), while some can cause formatting issues (zero-width space). 

Some of the most common Unicode characters that ChatGPT uses include:

Em Dash

The long dash character ChatGPT frequently inserts—especially in recent model versions (o3, 4 o, 4.1)—to break up sentences or add dramatic pauses.

• Unicode: U+2014
• Description: A long horizontal dash wider than a hyphen. Useful in typography for parenthetical breaks, but in code, CSV, or plain-text data it can act like a non-standard character, causing copy/paste or parsing issues.
• Example: word—word ← visually similar to word-word, yet represents a distinct Unicode code point.

Smart Quotes

The curly “typographer’s” quotation marks that ChatGPT sometimes substitutes for straight quotes.

• Unicode:
  
  • Left double quote (U+201C)
  • Right double quote (U+201D)
  • Left single quote (U+2018)
  • Right single quote (U+2019)
• Description: Curved punctuation intended for print typography. They often appear where a plain straight quote (", ') would be safer—breaking code snippets, CSV files, or Markdown.
• Example: “hello” or ‘world’

Zero-Width Space

A completely invisible spacing character that may slip into text when copying from ChatGPT or other editors.

• Unicode: U+200B
• Description: Adds no visible gap but still separates characters. Can break string matching, URLs, and word counts; causes “why won’t it paste correctly?” problems.
• Example: word​word ← looks like one word, but a zero-width space sits between the two halves.

Why do LLM’s Use These Hidden Characters?

There are 3 reasons that likely contribute to the increased usage of invisible characters by LLMs:

1. Training Data Bias: Since LLMs train heavily on professionally edited texts where em dashes and smart quotes are standard, it would have learned to use them more frequently than in everyday writing.
2. Mimics Formal Tone: LLMs have a bias to sounding formal/authoritative and these characters help it achieve that.
3. LLM’s Don't Use a Keyboard: The reason we don’t use these invisible characters more is because they are not on any standard keyboard (ASCI, ISO etc). LLMs don’t “write” using a keyboard so it is no extra effort for them to use an invisible character compared to a human writing.

Common Uses and Concerns with Invisible Characters

Why do people use them?

• Tidy text layout - A zero-width space (U+200B) or soft hyphen (U+00AD) lets writers nudge where a line breaks, so long words don’t dangle awkwardly at the edge of a column.
• “Spaces” in usernames - Some sites forbid real spaces. Sneaking in an invisible Hangul Filler (U+3164) keeps John Doe readable while still passing the “no-space” rule.
• Subtle watermarks - Publishers can hide a unique pattern of zero-width marks inside an article. If the text leaks, those invisible dots act like digital fingerprints.

Why do people worry about them?

• Formatting - Looks the same, acts differently - Two snippets that appear identical may hash or sort differently once hidden characters are factored in, breaking exports, searches, or audit trails.
• Invisible hiding places for bad code - Attackers can bury malware or secret instructions among zero-width characters; the file compiles or runs, but a human code reviewer sees nothing unusual.
  
  • Example: https://www.promptfoo.dev/blog/invisible-unicode-threats/
• Silent prompt tricks on AI - Hidden Unicode can smuggle extra instructions into a chatbot prompt, making the model reveal data or generate harmful content without the user noticing.
  
  • Example: https://www.prompt.security/blog/unicode-exploits-are-compromising-application-security
• Your Text Looks AI-generated - If text has a lot of unique formatting that is heavily used by AI (Em Dash etc) it is clear that it was AI-generated which can cause reputational harm.
  

Table of All Common “Invisible” Unicode Text Characters

Below is a complete list of the most common invisible text characters and how our free tool handles them.

‍

Does Adding or Removing Invisible Characters Help Bypass AI Detectors?

No - Our short test below shows that the addition or removal of hidden characters did not change the detectability of AI-generated content. We created 2 pieces of AI content and modified it by adding extra invisible characters, as well as stripping all Unicode characters using our free tool on this page.

The result was that the content was detectable by most tools regardless of the status of Invisible Characters. 

Below are the findings…

Text AI Sample 2 - Extra Invisible Characters had a TON of Hidden Characters!

Getting ChatGPT to intentionally add extra characters resulted in this warning from them…

Do All LLMs Inject Hidden Characters?

Yes - We looked at the same prompt in several popular LLMs to see if Hidden Characters are

Prompt: Write a LinkedIn post about the benefits of formatting LinkedIn posts. 

No model produced an Invisible Character but all relied heavily on formatting with visible Unicode characters. 

Example  ChatGPT 4o:

:

‍