Clean AI text
Paste the text: every fix is listed and can be undone, and typical AI phrasing is highlighted, with no verdict and no score.
Your files never leave your deviceOr drop a .txt or .md file
The file is read in your browser.
Why no “written by AI” score? AI detectors are often wrong, especially on texts by people writing in a language they learned. So VellumPDF shows no percentage and does not promise to “beat” a detector: the tool removes technical traces and shows you phrasing worth revising.
How to clean text copied from ChatGPT, step by step
The pasted text, or the .txt or .md file, is analysed in your browser: nothing is uploaded.
Paste the text or drop a file
Paste the answer from ChatGPT, Claude, Gemini or Copilot into “Text to clean”, or drop a .txt or .md file.
Click “Clean the text”
VellumPDF finds invisible characters, special spaces, look-alike letters, leftover Markdown, source markers, link trackers and chatbot phrases.
Check the fixes
Under “Technical fixes”, each group can be expanded. Untick a line to undo that fix: the cleaned text updates at once.
Reread the highlighted hints
“Style hints” flag phrasing that is common in AI output. Rephrase the ones that don’t sound like you: they are not mistakes.
Copy or download the result
“Copy text” puts it on the clipboard; “Download (.txt)” saves it as a text file.
The tool does not say whether a text was written by AI and does not promise to fool a detector: it removes technical traces and shows phrasing worth revising.
Frequently asked questions
Does the tool tell whether a text was written by AI?
No. It gives no score and no percentage. AI detectors are often wrong, especially on texts by people writing in a language they learned, and VellumPDF does not show a number it cannot check. The tool removes technical traces and highlights “Style hints”: phrasing that is common in AI output and that people use too.
Which invisible characters are removed?
Zero-width spaces (U+200B), the word joiner (U+2060), the byte order mark (U+FEFF), the soft hyphen (U+00AD), direction marks (U+200E, U+200F, U+202A to U+202E, U+2066 to U+2069), Unicode tag characters, which can carry hidden text, and a few other zero-width characters. The joiners U+200C and U+200D are only removed between Latin letters: they stay in combined emoji and in Persian or Hindi. Direction marks stay in Arabic or Hebrew text.
Are French non-breaking spaces kept?
Yes. A non-breaking space stays before a colon, a semicolon, an exclamation or question mark and a closing guillemet, after an opening guillemet, and next to a digit (1 250 francs, 10 km). Anywhere else it becomes a normal space. Ideographic spaces stay in Chinese, Japanese or Korean text.
What happens to Markdown formatting?
Heading hashes, bold and italic asterisks, code backticks, quote markers, horizontal rules and code-block fences are removed. “- ” bullets become “• ”, and a [text](address) link becomes “text (address)”. Markdown tables are left as they are.
Which chatbot phrases are removed?
At the start, a line such as “Sure! Here’s a polished version:”. At the end, an offer from the assistant such as “Would you like me to shorten it?” or “Let me know if you need anything else”. A closing line addressed to the recipient of a letter is kept. These phrases are recognised in French, German, Italian and English. If a removal looks wrong, untick it.
Is everything the chatbot adds removed?
No. Image captions in an answer with web search, for example, look like ordinary text and stay in place. The tool lists what it changed, not what it may have missed: always reread the cleaned text.
Is the text sent to a server?
No. The analysis and the fixes run in your browser: the pasted text or the dropped file never leaves your device. “Copy text” goes through your system clipboard, and “Download (.txt)” saves a UTF-8 text file.