Unwatermark
Open the app: Clean AI watermarks from text → (products.documentize.ai)
In one sentence
Unwatermark removes invisible markers hidden in AI-generated text (and the file traces that point to an AI tool) from a text file or a PDF's text — it does not erase visible watermarks such as "DRAFT" stamps, logos or page backgrounds.
Read this first: what it can and cannot do
The name suggests one thing; the app does another. The code and the site describe a text cleaner for AI watermarks:
| It can | It cannot |
|---|---|
| Find and remove hidden statistical watermarks in AI-written text | Remove a visible stamp, logo, "CONFIDENTIAL" banner or background image from a PDF page |
| Strip provenance traces: metadata, hidden tags, a vendor name in the file name (optional) | Restore the original look of a PDF — what comes out is plain text laid out afresh |
| Score how AI-generated a text reads (detect-only mode) | Prove a text is or is not AI-written — the score is an estimate |
| Rewrite the prose so the statistical pattern is gone, while keeping facts, numbers and names | Keep your layout, images, tables, fonts or links |
For visible marks on a page, there is no automatic eraser in Documentize. Honest options: go back to the source and export without the mark; ask whoever added it for a clean copy; or cover your own content with Redactor (that covers, it does not recover what lies under the mark). Removing someone else's watermark may also breach their rights; that is your responsibility.
What makes it unique
- It is the only app in the book that works on the statistical fingerprint of AI-written text, not on pages. Look-alikes: Redactor hides visible content, Recovery sanitizes a PDF's structure and metadata. Unwatermark deals with the wording itself.
- Detect-only mode gives a report with no file, so you can check before changing anything.
- It gives you a before/after report showing what changed, and highlights the changes in the pasted-text view.
- You can paste text straight into the page instead of uploading a file; you get it back to copy, and as a PDF.
When to use it / when not to
- Use it to check or clean text you generated yourself with an AI tool before publishing, or to remove provenance traces from a text file.
- Not for watermarks you can see (see the table above).
- Not for Word, ODT or other office files. The site's marketing mentions DOCX and ODT, but the processing code accepts plain text files (such as .txt, .md, .csv) and PDF, and rejects other binary files with a clear message (verify on the live app). Convert to PDF or text first with Converter.
- Not for documents whose layout matters. A PDF comes back as plain re-typeset text.
Before you start
- Only one file per run.
- A PDF must have real text. A scanned PDF has none: run Searchable first.
- Decide whether you accept rewriting. It is on by default and changes your wording (slower, but it checks that facts, numbers and names are kept).
- Read the result. Rewriting by a model can change nuance.
Step by step
- Open the Unwatermark page and either drop a file or paste text into the box.
- Choose your options (below).
- Press the button and wait ("Removing watermarks…", or "Detecting AI-generated text…").
- Download the cleaned text, the PDF version and the report — or read the detection report on the page.
Options explained
| Option | What it does | Default | Recommended |
|---|---|---|---|
| Remove AI vendor names from the file name | Strips an AI tool's name from the output file's name | off | On if the file name gives away the source |
| Rewrite the text with an AI model | Rephrases the prose in addition to the deterministic cleanup; checks that facts, numbers and names are kept; slower | on | On to defeat statistical watermarks; off for a fast, wording-preserving cleanup of hidden characters and metadata |
| Only check whether the text is AI-generated | Report only; the file is not changed and nothing is produced for download | off | Use it first when you only need an opinion |
| Detection threshold (0-100) | AI score at or above which the text is reported as AI-generated; lower is stricter | 35 | Keep 35 unless you have a reason. The code's note: ordinary human text scores around 15, strongly AI-styled text around 50 |
What you get
- For a text file: the cleaned text file, plus a report.
- For a PDF (or pasted text): the cleaned text, a new PDF with that text laid out again (without the original's metadata), and the report. The report compares the original and cleaned PDF's author, creator, producer, title and keywords and lists the page counts.
- In detect-only mode: a report on the page, no download.
- Your originals are never changed.
Quality & limits
- PDF layout is lost. Images, tables, columns, fonts and links are gone; only text in reading order remains, set in one simple style, so the page count may change.
- Detection is probabilistic. The site claims over 95% detection accuracy on common watermarking techniques; treat that as the site's claim. Scores can be wrong both ways.
- Rewriting can alter meaning. The site says only watermarked portions are rewritten; the page's own hint says it rephrases prose while checking facts, numbers and names. Always read the output before use.
- Text only. Non-text files are rejected.
- Very long texts: the page shows the before/after view only up to a size limit; beyond that you still get the downloads.
- One file per run. No continuation is offered after a free-plan time stop.
- The site marks the feature as free during beta (verify on the live app). Results are kept for about two days.
Works well with
- Before: Searchable (if your PDF is a scan), Converter (office file → PDF).
- After: Converter (back to Word, if you need), Add TOC, Protector.
- Related, not the same: Recovery (structural sanitizing), Redactor (visible content).
Common mistakes & fixes
- "My 'DRAFT' stamp is still there" → this app does not touch visible marks (see above).
- "My formatted report became a plain page of text" → by design for PDFs. Keep your original for the formatted version.
- "It says the file is binary / not supported" → convert it to PDF or text first.
- "Empty result from a scanned PDF" → there was no text to extract; OCR it first.
- "The wording changed and I did not want that" → turn off Rewrite the text with an AI model.
Same thing elsewhere
- CLI plugin: Documentize.Unwatermark — a local, deterministic cleaner for hidden characters (zero-width and tag characters, odd spaces, look-alike letters). Its options are detect only and normalize punctuation (typographic quotes, dashes and ellipses become plain ASCII). It has no AI rewriting, which is a difference from the web app. See Beyond the browser.
- MCP: no dedicated tool was verified.
Try it
Paste a few paragraphs of text you wrote yourself into the box and run it with Only check whether the text is AI-generated. Note the score. Run it again on a paragraph produced by an AI chat tool and compare. Then run a cleanup on the latter with Rewrite off, and read the report to see what (if anything) was found.