Precision

A computer program designed to detect something keeps reporting hits. A spam filter, for example, flags: This email is junk mail. Precision answers the question of how often such reports are correct. To do this, one counts all reported hits and checks how many of them were actually correct. If the filter flags 100 emails as spam and 90 of them are truly spam, the precision is 90 percent. The value always lies between 0 and 1, often written as a percentage.

What a false alarm costs

Low precision means many false alarms. For a spam filter, this means: important emails end up in the trash folder. For a bank whose software flags suspicious transfers, it means: harmless customers get blocked and have to call in. Each of these cases costs time, money, or trust. That’s why precision is often the metric companies scrutinize most closely.

A single figure like “95 percent correct,” by contrast, says surprisingly little. Imagine a disease that one in a thousand people has. A program that simply says “healthy” for everyone is correct in 99.9 percent of cases. Yet it is completely useless. Precision forces you to look only at the reported cases, and it exposes such illusory successes.

It’s important to distinguish precision from recall, its counterpart. Recall doesn’t ask about the reports, but about the actual cases: how many of them did the system even find? A system can have perfect precision and still be bad if it reports almost nothing.

The math behind the metric

A system’s reports are divided into two groups. The correct hits are called True Positives — that is, detected spam that really was spam. The false alarms are called False Positives — that is, flagged emails that are actually harmless. Precision is the number of correct hits divided by all reports combined. Cases the system never reported at all do not appear in this calculation.

In practice, precision can almost always be shifted after the fact. Most models don’t output a hard yes-no answer, but a confidence value between 0 and 1. Only a threshold turns this into a decision. If you set the threshold high, the system only flags very clear-cut cases. Precision rises, but many real cases slip through.

This is exactly where the trade-off lies. Precision and recall can rarely be maximized at the same time. Anyone wanting to combine both into a single figure uses the F1 score, a kind of average of the two. Which side matters more is decided not by mathematics, but by the use case.

Precision in products and headlines

In news about AI systems, precision usually comes up when a vendor wants to prove its detection performance. Makers of credit card fraud software, medical image analysis, or text-checking tools cite the figure in data sheets. It becomes problematic when only a single value is given. Without stating recall and the test data, a high precision is easy to produce.

A well-known example is programs meant to detect whether an essay was written by an AI. Even with decent precision, false alarms remain, and each one hits a student who wrote the text themselves. That’s why several providers have since discontinued such tools. The metric alone says nothing about how severe an error actually is.

You also encounter this principle constantly in everyday life without hearing its name. A phone’s facial recognition, a keyboard’s autocorrect, and a streaming service’s recommendations are all evaluated by the same pattern. Anyone who reads a hit rate in the future should ask the follow-up question: How many reports were there in total, and what was missed in the process?