False positive rate

Also called FPR

The false positive rate is the proportion of human-written documents a detector wrongly flags as AI-generated — the single number that decides whether a tool is safe to use for accusations, and the one most often quoted without its denominator.

Last reviewed 15 August 2026 · The HumanFlow team

In plain English

Run a detector over a hundred essays that people definitely wrote themselves. However many come back flagged, that share is the false positive rate.

The reason it matters more than accuracy: accuracy can be high while the false positive rate is still catastrophic at scale, because most submitted work is not AI-generated.

A worked example

Why a 1% false positive rate is not the reassurance it sounds like. The arithmetic below uses a stated rate against a stated volume — no measurement of our own.

Turnitin stated it had reviewed more than 200 million papers
for AI writing (company press release, 2024).

At a 1% false positive rate:   200,000,000 x 0.01 = 2,000,000
At a 0.1% false positive rate: 200,000,000 x 0.001 =   200,000

Two million wrongly flagged documents, at the rate that sounds like a rounding error. Even at a tenth of that, two hundred thousand.

The number attached to any individual student is not 1%. It is the probability that this specific paper is one of the wrongly flagged ones, and that depends on how much AI writing is actually in the pile — which nobody knows.

This is arithmetic on a vendor's own figures, not a measurement of any detector's real rate. The point is the shape of the problem, not a claim about a particular tool.

Why it matters for AI detection

It is the number to ask for when an institution proposes using a detector for discipline, and the number vendors are least forthcoming about — usually stated for a whole-document verdict on a favourable corpus.

It also varies by writer. A single headline rate averages over populations that are affected very differently; the same tool can be near-harmless for one group and unreliable for another.

If you are appealing a decision, the rate is the argument. Not "detectors are sometimes wrong" but "here is the published rate, here is the volume it is applied at, and here is what that produces."

Commonly confused with

False positive: One wrongly flagged document, versus the proportion across a set.
Accuracy: Accuracy counts all correct answers together. A detector that flags nothing is 95% accurate on a corpus that is 95% human — and useless. False positive rate is the half of the picture that decides who gets accused.

Sources

Turnitin — AI detection feature reviews more than 200 million papers

Part of the AI detection glossary.