False negative

In plain English

The detector says human. The truth is machine. Nobody appeals a false negative, so the only errors an institution ever hears about are the other kind.

That asymmetry in feedback is worth naming: a tool can look reliable in practice simply because half its mistakes are invisible.

A worked example

The four outcomes, and which of them anyone ever finds out about.

                    detector says AI    detector says human
  actually AI       true positive       FALSE NEGATIVE
  actually human    FALSE POSITIVE      true negative

  reported to the institution:  false positives only
  discovered by anyone:         false positives only

A student wrongly flagged appeals, and the error enters the record. A student whose AI-written essay passed says nothing, and the error never enters anything.

So an institution reviewing its detector's performance is reviewing a sample of errors that is systematically incomplete, in one direction.

Weber-Wulff and colleagues measured this end of the problem directly: across fourteen detectors, accuracy on machine-paraphrased text fell to 26%. Most AI text put through a paraphraser was missed.

Why it matters for AI detection

Because accuracy claims are meaningless without both error types, and vendors have every incentive to foreground the one that makes them look careful.

It also undercuts the main argument for using these tools at all. If a detector misses most paraphrased AI text, screening mainly catches people who did not try to hide — which is not the population an integrity process is aimed at.

For anyone weighing policy: a tool with a very low false positive rate and a very high false negative rate is not a cautious tool. It is a tool that has moved its threshold until it rarely says anything.

Commonly confused with

False positive

A false positive accuses the innocent; a false negative excuses the guilty. Detectors trade one against the other, and no threshold eliminates both.

Detection threshold

The threshold is the dial that trades these two errors against each other. Neither rate is a fixed property of a detector.

Read next

• How accurate AI detectors are
• False positive
• The accuracy research, reviewed