Upload a PDF up to 10 MB and every page goes through the detector on its own, giving you a page-by-page map of exactly where the AI-looking writing sits plus a document score on top. Word .docx and .txt are accepted too, but they carry no fixed page boundaries, so those return a document score only. The first 3 documents a day need no account and no card. There is no OCR here, so a scanned PDF needs OCR-processing first.
A clean introduction followed by three pasted pages averages out to a middling document score, which tells a reviewer nothing they can act on. Scored page by page, the same file stops hiding.
Illustration of the output shape. This document averages to a forgettable 52 overall; page by page it points straight at pages 4 to 6. Page 8 is a reference list under 30 words, so it is reported unscored rather than given a number.
A PDF is split at its real page boundaries first, then every page is sent through the same classifier independently. Nothing is averaged before scoring, so a spike on one page cannot be diluted by twenty clean ones around it. Pages are processed a few at a time against a bounded budget, which is why a long document returns "scored the first N pages" rather than timing out.
This is the one honest caveat on the format list. A PDF carries fixed page boundaries, so the map lines up with the pages you can see. A .docx or .txt has no fixed pages — the page breaks you see in Word are produced at render time by your font and margin settings — so those files are accepted and scored, but they return a document score without a page map. If you want the map, export the file to PDF first, which takes one menu click.
Every scored page returns 0 to 100 with a band attached: 70 and above reads as AI, 40 to 69 as mixed, below 40 as human. The bands exist so a reviewer scanning a 40-page map can find the red pages without reading forty numbers. The underlying score is always shown next to the band.
Pages under roughly 30 words are reported as unscored rather than given a number. Title pages, reference lists, section dividers and stub final pages give the detector almost nothing to work with, and a confident-looking score on twelve words is noise, not evidence. The page still appears in the map so the count adds up.
No account for the first three documents a day, no extension to install, no copy-paste through a text editor.
Open the document detector and drag a PDF, .docx or .txt onto the upload area, or click to pick it from a folder. The cap is 10 MB per file. The file type is checked by its actual contents rather than its extension, so a spreadsheet renamed to .docx is caught and refused with a message that explains why.
PDFs are read for their selectable text layer with page boundaries intact, so the page numbers in the result line up with the page numbers in the file you uploaded. Word files are read out of the document XML, which has no fixed pages, so a .docx is scored as one document. Nothing is retyped, reflowed or pasted through an intermediate textarea.
Each page goes through the detector separately and comes back with its own 0 to 100 score and band, alongside a document-level summary. Pages past your tier's scoring cap are still counted and reported in the page total; they simply do not each get an individual score.
This tool reads prose. Being specific about what it turns away matters more than a long format list, because a refusal with a reason is more useful than a scan that silently returns nothing.
Text-extractable PDFs, Word .docx files, and plain .txt, up to 10 MB each. These cover the overwhelming majority of what actually gets submitted: exported essays and theses, chapters, reports, contracts, proposals, and anything saved out of Word or Google Docs as a PDF or .docx. Only PDFs come back with a page map; .docx and .txt return a document score, because neither format has fixed pages to split on.
Spreadsheets (.xlsx, .xls, .csv), slide decks (.pptx, .ppt), legacy .doc, and .pages, .odt, .rtf and .epub are turned away with a message telling you what to do instead, usually exporting to PDF or .docx. Because .docx, .xlsx and .pptx are the same underlying container, a workbook or deck renamed to .docx is caught by inspecting the contents rather than trusting the extension.
A PDF from a flatbed scanner or built from phone photos holds pixels, not characters, and there is no OCR step here. Run the scan through any OCR tool first and upload the searchable PDF it produces. Saying so plainly is better than claiming an OCR layer that does not exist.
The document detector takes one file at a time and returns a page map for it. If you need to queue many documents at once and sort the results by score, that is Bulk Scan on the Business tier, which is a different surface with a different workflow.
Two separate numbers, and they are worth keeping apart. One is how many files you can scan in a day; the other is how many pages inside each file get their own individual score.
Enough to check a real submission before deciding whether the tool is worth an account. This anonymous allowance is shared with the text AI detector, so three document scans and three text scans come out of the same daily pool of three, not six.
Covers a full essay or a thesis chapter end to end rather than just the opening pages. No card required.
Sized for a student or a light reviewer working through submissions a few times a week.
Pro scores 100 pages per document and allows 100 documents a day. Business and Enterprise lift the daily document limit while keeping the same 100-page scoring depth per file. The page caps are latency limits rather than billing limits: pages past the cap are still counted and reported, they just are not each scored individually.
Start with no account at all. Paid tiers billed in USD with yearly billing saving 25%. Full details on the pricing page.
Billed $89.88/year, Save $30
Billed $179.88/year, Save $60
Billed $359.88/year, Save $120
Four situations where knowing which page matters more than knowing the average.
Long documents are rarely uniform. A chapter drafted across a semester can be entirely the student's own except for a literature summary pulled from a chatbot in a bad week. A document-level score buries that; the page map puts it on one line.
"The document scored 58" is not a conversation anyone can have. "Pages 4 through 6 scored 84 to 91, the rest sit under 35" is specific enough to open the file and read those pages together. Detectors are not proof, and the page reference is what makes the follow-up conversation about the writing rather than about the tool.
Multi-author documents are the natural case for per-page scoring, because contributions usually break along section boundaries. The map shows which contributor's section is worth a second look before the document goes out.
Copy-pasting a 30-page PDF into a text box collapses paragraphs, drags in running headers and page numbers, and takes long enough that most people simply do not bother. Uploading the file keeps the page boundaries the result refers to.
A per-page map is better evidence than a single number. It is still not proof, and we would rather say so here than have you find out from a bad decision.
AI detectors work on statistical patterns in writing. Careful, formal, heavily-edited prose can read as machine-generated, and lightly-edited AI output can slip under. The right response to a red page is to open it and read it.
Writing produced by someone taught English formally as a second language shares measurable surface features with AI-generated text: consistent sentence length, conventional connectives, restrained vocabulary. This is a documented weakness of detectors generally, including this one. Weigh a flagged page against what you know about the writer.
The classifier is trained and calibrated on English. Documents in other languages will return a score, but it is not a number to rely on.
The detector reports how AI-like the writing reads. It does not identify which assistant produced it, and any tool claiming to name the specific model from prose alone is overstating what is possible.
Most tools in this category accept a file and return one number. These are the questions that separate them, with our own answers stated plainly so you can hold us to the same standard.
Many detectors accept a file upload but then truncate the extracted text to whatever their paste box accepts, scoring only the opening pages while presenting the result as a verdict on the whole document. Ask what happens to page 30. Our answer: pages are scored individually up to your tier's cap of 5, 20, 50 or 100, and pages past the cap are reported as counted-but-unscored rather than quietly dropped.
A single document percentage cannot be acted on. If a tool says 58% without saying which pages, you still have to read the whole file. Our answer: every scored page carries its own number and band, so the map points at the pages worth opening.
A title page or a reference list has almost no prose in it. A detector that returns a confident score on twelve words is inventing a number. Our answer: anything under roughly 30 words comes back labelled unscored, and it still appears in the page count so the totals add up.
Plenty of tools imply that any PDF will work. A scanned PDF holds pixels, not characters, and without an OCR step there is nothing to score. Our answer: there is no OCR here, we say so before you upload, and the fix is to run the file through any OCR tool first.
Free tiers in this category often mean a handful of lifetime uses, or a result hidden behind a signup or a card. Our answer: 3 documents a day with no account, no card and no email, sharing one pool with the text detector. If it is useful you will know before you have given us anything.
Where this tool is weaker: no OCR, no bulk queue, no spreadsheet or slide-deck support, and English-only calibration. If you need to sweep two hundred files at once, that is a different surface. If you need one document read properly, page by page, this is built for that.
The PDF-specific path, format honesty on text-extractable versus scanned files, and where extraction struggles.
Read the guide →Scanning .docx drafts: how Word files are read, and what survives the trip out of the document XML.
Read the guide →The tool itself. Drop a PDF, DOCX or TXT in and read the page map. No signup for the first three a day.
Open the tool →The general text detector, the paste-first flow, model coverage, and how the classifier scores prose.
Read the deep dive →PDF, DOCX or TXT up to 10 MB. A score for each page, not one number for the file. First three a day need no account and no card.