Batch scanning

What batch scanning is for

One pass over a set of documents, producing a table: filename, score, verdict, and how stable each result is. It exists to save the mechanical part of looking at thirty submissions - not to make a decision for you about any of them.

The useful way to read the output is as triage. It tells you which four or five documents are worth opening and reading closely. That is a real saving on a marking session, and it is honest work for the tool, as long as reading them is what happens next.

The output is deliberately shallow: a score and a stability reading per document, no highlights, no per-sentence attribution. For a document you actually care about, open it in the single-document scanner, which shows the passages, the measurements behind each flagged sentence, and produces a PDF report.

The arithmetic of a class set

A vendor quoting a 1% false-positive rate sounds like a rounding error. Across a cohort it is not, and the reason is base rates rather than accuracy.

Take a detector catching 80% of genuine AI use at a 1% false-positive rate - generous on both counts. In a class of 120 where 30% used AI undeclared, about 3% of your flags are wrong. Where 10% did, it is 10%. Where 3% did, it is 29%. Where 1% did, more than half the students you flag did nothing.

The consequence is uncomfortable and worth sitting with: the better your teaching, your assignment design and your classroom culture work, the lower the real rate of AI use - and the less trustworthy each individual flag becomes. A tool that is least reliable exactly when you are succeeding cannot be the thing that decides an allegation.

This is why the table shows a stability reading beside each score. A verdict that moves when the text is reworded in harmless ways is not measuring the writing; it is measuring one particular arrangement of characters.

What a ranked list does to a reader

Sorting a class by AI score produces something that looks like a finding and is not. The order is real; the meaning attached to it is not, because the differences between adjacent rows are usually smaller than the error on any single row.

The pattern to watch for in your own results is who ends up at the top. Detectors over-flag writing in an acquired language, prose that follows a taught essay template, formulaic genres like methods sections and lab reports, and heavily revised drafts. If your flagged names skew toward your international students, the tool has told you about itself rather than about them.

Nothing in the CSV export is evidence of misconduct. It is a list of measurements, and it should never be forwarded to a student, a colleague or a panel as though it were more than that.

What to do with a high score

Treat it as a reason to look, never as a finding. Open the document in the single-document scanner and read the passages it flags; a run of consecutive flagged sentences means considerably more than the same number scattered through a text.

Then look at the things a detector cannot see. Does the document cite sources that exist and say what it claims they say? Does it read like this student's other work? Can they talk about the draft - what they cut, what they struggled with, where the argument changed?

Version history in Google Docs or Word settles most cases faster than any score, because a document that accumulated over days with false starts and restructuring is very hard to fabricate after the fact. Ask for process, not for a confession.

How it works in practice

Up to 40 documents at a time, as .docx, .pdf, .txt or .md. Each file is read in your browser and scanned one after another rather than all at once - parallel requests would queue behind each other anyway on a single machine, while looking a great deal like an attack.

A class set takes a few minutes. Progress is shown per file, and a document that fails to read - a scanned PDF with no text layer, for instance - is reported and skipped rather than stopping the run.

Nothing is uploaded and nothing is kept. Extraction happens on your device, only the extracted text is sent to be scored, and closing the tab discards the results. Export the CSV before you leave if you want them.

Common questions