For text copied out of ChatGPT, Claude and friends
Looks tidied by hand.
Unbot removes invisible AI markers you cannot see, but your professor might.
- 1. Paste or upload your text.
- 2. Optional: choose your font and cleaning rules.
- 3. Click Clean My Text.
- 4. Copy the result or download it as .docx.
01 · Original
Paste or upload
02 · Result
Clean copy
Your cleaned text shows up here.
Why cleaning matters
Rules from research, not guesswork.
01. INVISIBLE CHARACTERS
Bad Characters: Imperceptible NLP AttacksBoucher et al., University of Cambridge, IEEE Symposium on Security and Privacy, 2022
Showed that invisible characters, look-alike letters and direction marks change how software reads text, even when it looks the same, and that text should be cleaned of them.
02. HIDDEN SPACES IN CHATGPTIndustry report
New ChatGPT Models Seem to Leave Watermarks on TextRumi, 2025
Found OpenAI's o3 and o4-mini models adding hidden narrow no-break spaces to long answers. OpenAI called it a training quirk, not a watermark.
03. MARKDOWN, LISTS AND EMOJIS
From Lists to Emojis: How Format Bias Affects Model AlignmentZhang et al., ACL, 2025
Found that the systems used to train AI favor lists, bold text, links and emojis, so AI models lean on them.
04. EM DASHESPreprint
The Last Fingerprint: How Markdown Training Shapes LLM ProseFreeburg, arXiv, 2026
Linked heavy em dash use in AI writing to markdown in its training data. GPT-4.1 used about 3 times more em dashes than people in essays.
05. CHATGPT LINKSNews report
Detecting ChatGPT in Legal Documents Is Easier Than You ThinkBloomberg Law
Found court opinions citing links that end in "utm_source=chatgpt.com", a tag ChatGPT adds to links.
06. FANCY LETTERSStandard
Unicode Normalization Forms (Unicode Standard Annex #15)Unicode Consortium
Sets the standard way to turn styled, full-width and joined letters back into normal letters.
Feedback
Found a bug or have an idea? Tell me.