Detection threshold

Also called Cutoff · Reporting threshold

A detection threshold is the score above which a tool reports a document as AI-written — a line the vendor chooses, which means the same text can pass or fail depending on where the line was drawn.

Last reviewed 15 August 2026 · The HumanFlow team

In plain English

Every detector produces a continuous number and then has to turn it into a verdict. The threshold is where it decides to stop saying "probably fine" and start saying "flagged".

Move the line down and you catch more AI text and accuse more innocent people. Move it up and the reverse. There is no setting that avoids both, which is the whole difficulty.

A worked example

One unchanged document, three thresholds. Nothing about the writing differs between the rows — only the line.

  document score: 18%

  threshold at 10%   →  FLAGGED    (report generated, case opened)
  threshold at 20%   →  suppressed (Turnitin shows an asterisk here)
  threshold at 50%   →  clean      (nothing surfaces at all)

The score never moved. The student never rewrote anything. Whether this document produces a misconduct meeting is decided entirely by a number set in a vendor's configuration.

Turnitin's own policy sits at the middle row: it attributes no score and no highlights between 0% and 20%, showing an asterisk instead, because its testing found more false positives in that band. Above 20% it states it aims to keep false positives under 1%.

So the asterisk is a threshold artefact, not a result. Reading it as "clean" and reading it as "suspicious" are both wrong.

Why it matters for AI detection

It is the setting that converts a measurement into an accusation, and it is almost never disclosed to the person being accused.

It also explains why two institutions running the same tool on the same essay can reach different outcomes. The disagreement is not about the writing; it is about configuration.

If you are challenging a finding, the threshold is a fair question to ask: what is it set to, who chose it, and what false positive rate did the vendor publish at that setting?

Commonly confused with

AI writing score: The score is what the tool measured. The threshold is the line it compares that measurement against. Same score, different threshold, opposite verdict.
False positive rate: The rate is a consequence of the threshold, not a fixed property of the tool. A vendor quoting a false positive rate without stating the threshold it holds at has not told you much.

Sources

Turnitin — AI writing detection capabilities FAQs

Part of the AI detection glossary.