AI detection for documents

Why a whole document needs more than a paragraph check

Most AI checkers expect you to paste in a block of text. That works for a single answer, but it falls apart the moment you're holding a 12-page essay, a 40-page report or a chapter draft. You end up copying sections one at a time, losing track of which slice came from where, and missing the part that actually matters: where in the document the writing shifts from a person's voice to a model's.

The Document AI Detector reads the file as a whole. It extracts the text from your DOCX or PDF, keeps the page boundaries intact, and scores each page on its own — so instead of one flat number for the entire submission, you get a map. A clean opening followed by three suspiciously uniform middle pages tells you far more than a single average ever could, and it points you straight to the passages worth a closer human read.

From file to per-page score

When you drop in a file, it's sent securely over HTTPS to the TextSight API. The first job is extraction: pulling the readable text out of the document while preserving where each page starts and ends. DOCX files and PDFs with a real text layer come through cleanly. Scanned, image-only PDFs have no selectable text to extract, so they need OCR first — that step lives on Pro.

From there, the same language model that powers our core detector runs across the extracted text. It weighs the statistical patterns that tend to separate machine-written prose from human writing — how predictable each word is given the ones before it, how evenly sentences are paced, and how little the rhythm varies from line to line. Human drafts are messier: they wander, hedge, repeat a favourite phrase, and break their own pacing. Generated text is usually smoother and more uniform than a person writing under deadline ever is.

Those signals are scored sentence by sentence, then rolled up into a number for each page and a verdict for the document. The riskiest sentences stay highlighted so the page-level score isn't a black box — you can see the exact lines that drove it.

Built for people grading and editing real files

Paragraph checkers are fine for a quick gut-check. This tool is for the people who get handed finished documents and have to make a call on them:

• Teachers & lecturers working through a stack of submitted essays without retyping each one into a box.
• Academic integrity reviewers who need to point to the specific pages and sentences behind a concern, not just wave at a percentage.
• Editors & content managers screening drafts from freelancers before they go into a publication or knowledge base.
• Hiring teams reading take-home assignments and written tasks that arrive as a finished document.
• Grant and proposal reviewers skimming long applications for sections that read as generated rather than written.

What the page scores are actually telling you

A per-page score is an estimate of how AI-like that page reads — not a confession and not proof. The most useful thing on the screen often isn't the headline number; it's the shape of the scores across the document. A draft where the writer leaned on a model for a tricky middle section will usually show a spike on those pages and lower numbers around them. A document that's uniformly high, or uniformly low, is a different story than one that swings.

Treat the result as triage. Let it tell you which pages deserve a careful read, then judge those pages yourself — checking whether the voice matches the rest of the work, whether the claims hold up, and whether the writer can talk through what they wrote. For anything with real consequences, the score should start a conversation, never end one.

Where it's strong, and where to be careful

No AI detector is infallible, and document-level checking has its own quirks worth knowing. Very short pages — a title page, a half-finished page, a references list — give the model little to work with, so their scores carry less weight than a dense page of prose. Heavily edited text, where someone generated a draft and then rewrote it in their own words, can read as human because, by that point, it largely is. And tables, citations and boilerplate can nudge a score in either direction without telling you much about the author.

A couple of habits keep you on the right side of all this. Don't act on a single page in isolation — read the document's pattern. Don't treat a high score as a verdict against a person, especially a non-native writer, whose plainer phrasing can look more uniform than it is. And if a result matters, pair it with the things a detector can't see: the writer's process, their drafts, and a quick conversation. Used that way, per-page scoring is a fast, honest first pass — not the last word.