The AI Badge Did Not “Miss” Your Rewrite—It May Be Scoring an Older Upload

People say the detector ignored their edits. More often, the detector never saw the edits.

Core point

An AI writing percentage is a classifier output for a specific submission snapshot. Your current .docx is allowed to disagree with that snapshot until you put the new bytes through the same pipeline (or until you stop comparing them as if they were one object).

Where the wires cross

• LMS attempt history vs desktop autosave

• Forwarded report PDFs without generation clocks

• Same paper title, new file created for “clean formatting”

• Instructor view pinned to an earlier attachment

Each produces a social story (“it didn’t update”) from a mechanical story (“different input”).

Literacy from the separation of systems

Turnitin-facing materials keep similarity matching apart from AI writing estimation. Authorship features, when present, are another separation. Version mismatch sits on top of all three: you can be out of sync on the AI dial while similarity still reflects overlap on the scored upload, or the reverse.

Also remember:

• Qualifying prose is not “every character in the PDF.”

• Short scored versions stay coarse.

• *% bands exist partly to discourage fake precision at the low end—compare bands across versions carefully.

A script you can reuse

“I am looking at the AI writing report generated at [time] for [filename/attempt]. After that time I changed [list]. The highlights on that report point to [passages]. I am not treating similarity or authorship as the same question unless you want those opened too.”

That paragraph is dull. Dull is good in office hours.

Soft CTA only after scope is clear

When revision of flagged prose is already decided, import the report that matches the upload, rewrite only marked passages, preserve citations and formatting. HumanPen is built for that document-level job—without a promise about the next badge.