Checking a PDF or Word document for AI detection
How a file gets turned into text decides part of your score before any model sees it, and most tools will not tell you what they extracted.
Most detectors take pasted text. A document is not pasted text, it is a file that has to be converted first, and the conversion is where a surprising share of odd results comes from.
What extraction does to a paper
Two-column PDFs interleave. A naive extractor reads across the page rather than down the column, producing sentences stitched from two different ones. Those fragments score unpredictably, because they are not sentences.
References become prose. A bibliography is hundreds of highly regular, highly templated lines. Fed to a classifier as text, it is close to the most machine-like content in the document, and it can be a large fraction of the word count.
Equations become noise. Depending on how they were embedded, they come out as scattered symbols, as nothing, or as mangled words. Any of those distorts the sentence statistics around them.
Tables and captions splice. A caption pulled into the middle of a paragraph creates a sentence nobody wrote.
Headers and footers repeat. A running header extracted on every page adds dozens of identical short lines, which is repetition a detector reads as regularity.
None of these are detection failures. They are extraction failures that arrive at the detector looking like text.
Word files are cleaner, with two exceptions
.docx carries real structure, so extraction is generally reliable. Two things still catch people out.
Tracked changes and comments. Depending on the tool, these may or may not be included. A document extracted with every rejected edit inlined is not the document you wrote.
Footnotes. These get appended, inserted inline, or dropped. In a document where footnotes carry substantive text, that materially changes what was scored.
LaTeX is its own problem
If you paste .tex source into a detector, a large part of what gets scored is markup: preamble, environment names, citation keys. The resulting number is about your markup, not your prose. Full version in why LaTeX breaks AI detectors.
How to tell whether extraction went wrong
The tell is a score that does not match the text. A paragraph you know you wrote, scoring high, in a document with a lot of structure, is worth checking before you believe it.
The check: look at whether the tool shows you what it extracted. Most do not, which is the single most useful thing missing from most detectors. If yours does not, paste a clean paragraph in directly and compare it against the same paragraph's score inside the file. A large gap is extraction, not detection.
What to do about it
Score the prose, not the artifact. If the tool lets you exclude references, exclude them. If it does not, paste the body text and leave the bibliography out.
Prefer the source format over the PDF where you have it. A .docx extracts more reliably than a PDF built from the same file.
Check per passage, not per document. A document average over a badly extracted file is dominated by the extraction damage. Per-passage scores make the damaged regions visible immediately, because they read as anomalies rather than shifting the mean.
Run long documents whole. Length is the thing that makes a score reliable, and splitting a paper into chunks to fit a word cap costs you exactly that. See why detectors are unreliable on short text.
Ours takes .pdf, .docx and .tex directly with no word cap, and shows a score per paragraph so extraction damage is visible rather than baked into an average. It is free with no account.
Worth being precise about what it does and does not clean up, since that is the whole point of this page. On .tex it removes equation, table and bibliography environments and citation macros before scoring, so the number is about your prose. On .pdf and .docx it scores what the extractor returns, which means a trailing reference list is included. Per-paragraph scoring makes that easy to see and ignore, but the document number on a PDF with a long bibliography is pulled by it. Paste the body text if you want that number clean.