Precision and recall

In plain English

Precision answers "when it accuses someone, how often is it right?". Recall answers "of the people who did it, how many did it catch?".

A tool can be excellent at one and poor at the other, and the headline accuracy figure will not tell you which.

A worked example

One marking pile, one detector, three numbers. Ten of the hundred essays were genuinely AI-written; the detector flagged eight and was right about six of them.

  100 essays · 10 genuinely AI · detector flags 8, correct on 6

  precision  6 / 8   = 75%    when it flags, it is right 3 times in 4
  recall     6 / 10  = 60%    it caught 6 of the 10, and missed 4
  accuracy   94 / 100 = 94%   ← the number that gets quoted

  behind that 94%:  2 students wrongly accused
                    4 AI essays passed unnoticed

The 94% is arithmetically correct and practically worthless. It is high mainly because 90 of the 100 essays were human and the detector left most of them alone, which is not an achievement.

The two numbers that describe what happened are 75% and 60%: one student in four flagged was innocent, and two AI essays in five were missed.

This is why an accuracy figure quoted without precision and recall should be treated as marketing. It is the one summary statistic that improves as the problem gets rarer.

Why it matters for AI detection

It explains how a detector can be accurate and useless at the same time. Because most submitted work is not AI-generated, a tool that flags almost nothing scores well on accuracy while catching nothing.

It also names the trade-off behind every threshold decision. Lowering the line raises recall and lowers precision — more AI caught, more innocent people flagged. There is no setting that improves both, and choosing one is choosing who bears the error.

If a vendor publishes only accuracy, the useful question is which of the other two it is hiding. A high false positive rate hides in precision; a tool that misses paraphrased text hides in recall.

Commonly confused with

False positive rate

The false positive rate is calculated over all human documents; precision is calculated over flagged documents only. They move together but answer different questions.

False negative

Recall is what false negatives subtract from. A tool with 60% recall is one that produces false negatives 40% of the time.

Read next

• How accurate AI detectors are
• Detection threshold
• The accuracy research, reviewed
• The RAID benchmark