What's new
Explainer14 Nov 20265 min read

A 99% classifier can miss every case you care about

When only ten of a thousand cases need attention, labelling everything negative scores 99% accuracy and finds none of them, so the confusion matrix has to sit beside the headline score.

In a fictional queue of a thousand documents, only ten need a particular review, and an assistant is meant to flag them. A classifier that says “no review needed” every time gets 990 decisions right. Its 99% accuracy sounds reassuring, yet it detects nothing.

Google's classification metrics guide, read on 14 November 2026, defines accuracy, precision and recall and discusses imbalanced data. We use those definitions to build a small example. It claims no result for any real AI model.

Constructed predictions for 1,000 cases: ten positives and 990 negatives. No named model was evaluated.
Synthetic candidateTrue positivesFalse negativesFalse positivesTrue negativesAccuracy
Always negative010099099%
Candidate822097097.8%

The second candidate finds eight relevant documents and misses two. It also flags twenty irrelevant ones. Its accuracy is lower, at 97.8%, while its recall is 80%. Twenty-eight documents enter the review queue and eight are relevant, giving precision of about 28.6%.

Which mistake creates the work?

A false positive sends an irrelevant document for review. A false negative leaves a relevant document unflagged. How much each costs is part of the acceptance rule, which is simply the agreed test for whether a candidate is good enough. Overall accuracy cannot give that back once the counts have been thrown away.

The counts worth keeping are the four in the confusion matrix: true positives, false negatives, false positives and true negatives. The positive label also needs a plain definition. “Positive” means the condition you chose to detect, not necessarily a desirable outcome.

The always-negative candidate predicts no positives, so precision has a zero denominator. It should be shown as undefined or not applicable under whatever convention you state. Printing 100% would make a detector that detects nothing look perfect.

Review load belongs in the comparison

If a person has to inspect the flagged documents, the queue size and the observed review time go in the record. A candidate that finds more relevant cases can still demand more checking, and the decision should reflect the actual workflow.

The example leaves missed cases unpriced on purpose. A money value needs a stated assumption and a reason it fits the job. The acceptance rule is best settled with the people who handle the errors.

Where the classifier exposes a score and a threshold, the chosen threshold goes with the result, because changing it gives a different decision setup. Candidates should be compared at the conditions you would deploy, not at a threshold picked after seeing the test labels.

Does the test set look like the real queue?

The number of positive and negative examples in the set should be reported. A deliberately balanced diagnostic set can help inspect failure types, but its precision need not describe the natural incoming queue. Results from a diagnostic set and from a realistic mix are best labelled separately.

Rare positives that matter need enough reviewed examples to show their variety. A handful of near-duplicates does not support a claim of broad coverage. A held-back group, not used to improve the prompt, keeps the final check honest.

Ambiguous labels deserve a look before the model takes the blame. If reviewers disagree about whether a document qualifies, the rule needs clarifying and the decision needs an adjudication note. A reference label that keeps changing makes the comparison hard to reproduce.

Picking the headline number for the job

For a screening workflow, the natural lead is detected relevant cases and missed relevant cases, followed by false alarms and review load. Another task may suit a different order. Whichever headline you choose, the raw counts should stay available.

A large accuracy percentage tells you something only while it stays attached to the class mix and the errors behind it.