An AI classifier reports 99 percent accuracy. That sounds impressive until you learn that only one percent of the examples contain the event it is supposed to detect. Predicting “not present” every time would achieve the same accuracy while detecting none of those events.
This is a fictional arithmetic example, not a result from a real model. It shows why the distribution of cases belongs beside the headline score.
Name the two mistakes
A false positive flags something that is not actually positive. A false negative misses something that is. Their consequences depend on the task. In a routine document-sorting system, an unnecessary review and a missed urgent document may create very different costs.
Google’s classification-metrics explanation distinguishes accuracy, precision and recall. Precision asks how many flagged cases are truly positive; recall asks how many actual positives were found. A useful evaluation reports the underlying counts as well as ratios.
Read a small example before a large dashboard
Suppose a test has 100 documents, including 10 that need review. The model flags 8, of which 6 genuinely need review. It has 6 true positives, 2 false positives, 4 false negatives and 88 true negatives. Accuracy is 94 percent, precision is 75 percent and recall is 60 percent.
Those figures describe different parts of the same behaviour. Whether it is acceptable depends on what happens to the four missed documents and the two unnecessary reviews. No single percentage can make that decision for the organisation.
Use representative, appropriately separated test data and inspect errors by relevant subgroup or document type. Changing the classification threshold may trade one kind of mistake for another, so compare at the intended operating point.
Before selecting the model, write down which error is more consequential, who reviews uncertain cases and how performance will be checked after deployment. The score becomes useful when it supports that decision rather than replacing it.