AI PDF Checker: Was This PDF Written With AI?

What does an AI PDF checker check in your file?

An AI PDF checker reads the PDF file itself, not a pasted copy of its text. AnyFact’s AI PDF checker takes the text layer of every page, reads scanned pages with OCR, sends embedded images to its photo detector and lists the metadata signals the file carries. It then gives a verdict page by page: probably human, uncertain, probably AI or mixed.

That matters because a PDF is a container, not just text. One file can hold typed pages exported from Word or Google Docs, pages that are only pictures of paper, charts and photos, and a record of the software that produced it. An AI detector for PDF that only sees copied words misses most of that evidence.

PDFs are a large share of what people check. Of the first 112 documents analysed in the AnyFact beta, 50 were PowerPoint files, 31 Word files and 30 PDFs, and most of the people uploading them were students checking their own coursework.

How to check a PDF for AI (step by step)

Upload the original PDF, read the page-by-page result, then decide what the flagged pages need. These steps work for a report, an assignment, a thesis chapter or a PDF someone sent you.

1. 01Find the best version of the file. A PDF exported directly from the original document keeps its text layer and metadata; a PDF that was printed, scanned or re-saved by another app keeps less.
2. 02Check the limits: up to 20 MB and 50 pages. Split a longer file, for example by chapter, and check each part separately.
3. 03Upload the PDF to the document check. There is no text box and no copy-paste: an AI checker with PDF upload reads the file directly. One document analysis uses 3 credits, and a free account gets 10 a month.
4. 04Let the analysis run. The text is read page by page, scanned pages go through OCR, embedded images go through the photo detector and the metadata is read from the file.
5. 05Read the overall verdict, then the pages that drove it. A mixed result names the pages that look generated; an uncertain one says why the evidence is thin.
6. 06Open the signals and the explanation. They tell you whether the verdict comes from the text, from images, from metadata, or from several of them together.
7. 07Download the PDF report and keep it with your drafts and notes. If you believe the result is wrong, report it with the “the result is wrong” button.
8. 08Decide your next step from the pattern, not the headline. If you wrote the work, file the report with your evidence of process. If a section was written with help your course requires you to declare, declare it. If the PDF is someone else’s, use the flagged pages to prepare questions.

The explanation is written in plain language, in English, French or Arabic. Uploaded files are deleted after 30 days.

What does AnyFact read in a PDF?

It reads four things in a PDF: the text layer of each page, scanned pages through OCR, embedded images and the document metadata. Each part goes to the method built for it, and the results are brought together page by page.

When the text score falls in the grey zone, Claude, Anthropic’s model, reviews the content as a second opinion. It can move the score only a little and never decides alone. The explanation then says which parts of the file carried the verdict.

Long academic PDFs raise their own questions: an abstract, a methods section and a set of figures can each have a different history. Our AI detector for research papers explains how to read a thesis or a paper section by section.

Can you check a scanned PDF for AI?

Yes. A scanned PDF has no text to copy, so a paste-only checker cannot read it. AnyFact reads scanned pages with OCR (optical character recognition) in English, French and Arabic, for the first 20 scanned pages of the file, then scores the recognised text like any other text.

The quality of the scan decides the quality of the check. OCR turns a picture of a page into text, and every misread word changes what the text detector sees. A clean scan of a printed page gives clean text; a crooked phone photo of a page in shadow does not.

• Scan pages flat, straight and in even light, without the shadow of your hand or phone.
• Use one page per image rather than two pages side by side.
• Crop out the desk, the binder edges and anything that is not the page.
• If you still have the typed original, upload that instead: a real text layer is more reliable than OCR.
• Split a file with more than 20 scanned pages, so that every scanned page is read.

Handwriting is a hard case for OCR, and a handwritten page is rarely what an AI check is for: the page itself is better evidence of who wrote it. Pages in a language other than English, French or Arabic are outside what the OCR reads.

What can PDF metadata reveal about AI use?

PDF metadata can show which software made the file, which software wrote the PDF, and when. It can point to an AI tool, but it cannot show who wrote the words. In most PDFs, the creator field names the application the document was made in, and the producer field names the software that wrote the PDF itself.

Each of these is a signal, not proof: converters and “Save as” change metadata too. The result lists the signals it found with that caveat and reads them together with the text and image checks.

How to read a PDF AI detector result, page by page

Start with the pattern across pages, not with the overall label. The overall verdict sums up the document; the page-by-page view shows where the evidence is and what kind of evidence it is.

A mixed PDF is common and is not an accusation. Reports are often assembled from pieces with different histories: a title page from a template, an introduction drafted by hand, a reference list produced by a citation manager. The page view lets you judge each piece on its own terms.

Front matter, bibliographies, tables of contents and data appendices are predictable by nature. If the flagged pages are mostly these, focus on the pages that carry your argument.

Should you check the PDF or the original Word or PowerPoint file?

Check the original file when you have it, and the PDF when the PDF is what you submit or what you received. The original keeps structure that a PDF export can lose; the PDF is the exact version your reader sees.

Our AI detector for Word documents and AI detector for PowerPoint explain what each original format adds. Whichever you upload, a document analysis costs the same 3 credits.

What are the limits of an AI PDF checker?

An AI PDF checker gives an indication, not proof, and it works within fixed limits. Know them before you read too much into a result.

• Size: up to 50 pages and 20 MB per file. Split longer documents and check each part.
• OCR: only the first 20 scanned pages are read, in English, French or Arabic. Blurred, skewed or handwritten pages produce OCR errors that weaken the text check.
• Images: up to 30 per document; small images and thin strips such as icons, logos and rules are skipped. A PDF made mostly of images gives the text check little to work with.
• Short documents: under 300 words, the text check gives little evidence.
• Formats: PDF, DOCX and PPTX only. Legacy .doc and .ppt files are refused; save them as .docx, .pptx or PDF first.
• Translated, paraphrased and heavily edited text is harder to judge, and so is formal academic prose.
• Non-native writers are more exposed to false positives, because careful, plain second-language writing can look predictable to a model.
• Detectors disagree. Different tools use different models and thresholds, so the same PDF can get different verdicts on different sites.
• No detector, AnyFact included, predicts what Turnitin or your school’s own tool will say. They are separate systems.

Our guide are AI detectors accurate? explains why results differ between tools and how to read accuracy claims. If your school uses Turnitin, our Turnitin alternative page sets out what each tool is built for.

Checking your own PDF before you submit: what actually protects you

The best protection is evidence of how you wrote the work, not a clean detector result. A PDF is usually an export, so that evidence lives in the source: the Word file, the Google Doc or the LaTeX project you wrote in.

• Keep the source file and its version history: Word’s version history (File > Info > Version History, for files saved to OneDrive or SharePoint) or Google Docs’ (File > Version history > See version history).
• Keep your notes, outlines, reading list and early drafts, even rough ones.
• Keep your sources: links, copies of the articles you read, and your citation manager library if you use one.
• If your course allows AI tools for some tasks, keep a short note of what you used them for and declare it the way your course asks.
• Download the AnyFact PDF report and file it with the rest.

If a teacher questions your work, stay calm and ask what was flagged, by which tool and on which pages. Offer your drafts and version history, walk through how you built the argument, and explain your sources. Check your school’s academic integrity policy for how such cases are reviewed and how to appeal. A detector result, from any tool, is an indication and never proof of misconduct.