False Negative
What is a false negative?
A false negative is when a test or judge misses a problem that really exists. The check says everything is fine, but when you inspect the actual output, there is an error the check failed to flag. Like its counterpart the false positive, the term applies to any kind of evaluation, but it carries special weight when you use one LLM to judge the output of another.
In AI evals, a false negative means a real error went uncounted. A judge asked to find missed sub-groupings in a generated opportunity solution tree produces a false negative whenever it overlooks a grouping that a human grader would have caught. A judge that correctly identifies every instance where the error exists has zero false negatives.
Why do false negatives let real errors slip through?
A false negative is an error your measurement never sees. If the judge misses real errors, your error rate looks better than it is, and problems ship to customers uncounted. You may believe your AI output is improving when the judge has simply stopped noticing what's wrong—which makes false negatives harder to spot than false positives, because nothing in your metrics tells you they are there.
How do you catch false negatives?
Recall is the measure that catches false negatives. Recall measures the judge's accuracy when there is an error. Score the judge against a calibration set of manually labeled samples and see how often it flags the ones that genuinely contain the error. A judge with 100% recall catches every real error—zero false negatives. A judge with low recall is missing real errors, and its error counts understate the problem.
Related terms:
Learn more:
Last Updated: September 16, 2026