AI text and document cleaner
Inspect and clean selected invisible characters, labels or Office metadata, with a before/after preview.
Automatic AI-authorship verdicts are suspended. See our independent test results below.
How to clean text or an Office document
- Choose a file or paste text. Drop a DOCX, XLSX, PPTX, TXT or Markdown file, or choose Pasted text. The file limit is 25 MiB.
- Review the detected items. Inspect declared metadata and select what to remove. To find a visible label, enter the full paragraph, text box or cell.
- Preview and create a copy. Compare the before/after excerpts, then copy the cleaned text or download a new file. You can also save the analysis report. Open the copy to check it before sharing.
Use Try an example to inspect a sample with a zero-width space (U+200B) and a visible “Generated by ChatGPT” label.
What can be removed from each format?
- Word. Matching paragraphs and supported classic text watermarks.
- PowerPoint. Matching text boxes, including on slide masters.
- Excel. Matching text values and simple header or footer labels.
- Text and Markdown. Supported invisible characters and matching full lines. Files must be UTF-8.
What our tests actually showed
The initial experimental classifier performed well on its original corpus, but did not generalize sufficiently. We froze its weights and thresholds before this external test. Automatic verdicts are suspended; technical cleaning remains available.
| Reference texts | Test cases | High AI signals from the initial model |
|---|---|---|
| Historical French human prose · 3 authors | 36 | 0 |
| English human texts · RAID subset | 96 | 0 |
| English AI texts · RAID subset | 528 | 56 |
| New assistant-written French texts | 4 | 0 |
| New assistant-written English texts | 4 | 0 |
These small, selected samples do not establish a general accuracy rate. The French human references are historical prose, and the eight new AI examples come from one assistant. Controlled mixtures and formatting edits are documented separately in the report.
Sources and protocol: RAID · Maupassant · Verne · Bernard · Detailed evaluation report (JSON)
Frequently asked questions
Can the tool recognize AI-generated writing?
This version no longer provides automatic AI-authorship verdicts. Validation on an external corpus and newly written texts does not confirm the reliability of the initial model. The test results are published below. Declared labels and metadata are still inspected, without establishing authorship.
Does cleaning remove AI watermarks?
It removes only supported invisible characters, selected labels and metadata. It does not rewrite text or remove statistical watermarks embedded in word choices.
What does the before/after preview show?
It shows excerpts of the text values and metadata affected by your selection. Invisible characters are shown with their Unicode code. Long passages and large selections are abbreviated. This is not a Word, Excel or PowerPoint layout preview and it does not modify the original file.
What is included in the downloadable report?
The text report lists the filename, findings, declared information, current selection and, if a copy was produced, the items used for that copy. It contains neither the full text nor document excerpts. The report is created in your browser.
Which invisible characters can I remove?
The tool lists zero-width spaces, soft hyphens and several directional controls. Non-breaking spaces and emoji joiners are kept. Directional controls can be intentional in multilingual writing: preview the changes before removing them.
What do AI-related metadata fields mean?
A creator, generator or application field may name an AI provider. This is editable declared information and does not establish how the content was written. Select the supported fields you want to remove.
Are my files uploaded and is the original changed?
Reading, previews, reporting and cleaning take place in your browser. No document or text is uploaded for this processing. The original is kept and cleaning creates a separate copy.
Which formats are supported?
DOCX, PPTX, XLSX, TXT and Markdown are supported, up to25 MiB per file. Cleaning affects detected and selected items. PDFs, images, legacy formats, encrypted or signed files and macro-enabled documents are excluded. Text files must use UTF-8.