Recall

Recall is a metric for how completely a system finds the cases it is looking for. To calculate it, you first define what the sought case is: a spam email, a tumor on an X-ray, a matching document in a search. Then you count how many such cases existed in total and how many of them the system actually caught. Recall is the share of caught cases out of all existing ones. With 100 existing spam emails and 80 found, recall is 80 percent. The 20 overlooked emails are called false negatives: they would have been a hit, but were let through.

What a low recall costs in practice

The overlooked cases are often the most costly errors. In medicine, a low recall means that illnesses remain undetected. In payment fraud detection, it means that fraudulent transactions go through. In both cases, nobody notices right away that something is missing. That is exactly what makes this type of error dangerous: false alarms get noticed, oversights do not.

That is why accuracy alone is often not a sufficient measure. Suppose that out of 10,000 examinations, 50 are abnormal. A system that always says 'normal' is right in 99.5 percent of cases. Its recall is nevertheless zero, because it does not find a single real case. For rare events, recall is therefore the more meaningful figure.

However, recall has a sister value without which it means little: precision. It measures how many of the reported hits were actually genuine hits. A system that flags everything as suspicious has a recall of 100 percent and is completely useless. Both figures must always be considered together.

The threshold where everything is decided

Most AI systems do not output a hard decision, but a probability. A spam filter, for example, says: This email is 73 percent likely to be spam. Only a defined threshold turns this into a yes or no. If you set it low, say at 30 percent, a lot gets filtered out and recall rises. If you set it high, at 90 percent, more spam slips through.

This is exactly where the trade-off lies: a higher recall almost always comes at the expense of precision. Anyone who wants to catch more real cases also catches more innocent ones. Developers therefore deliberately set the threshold based on which type of error causes more harm. In cancer screening, a false alarm is acceptable, an overlooked tumor is not. With an email filter, it’s the other way around: a deleted job application is more annoying than a slipped-through promotional email.

To combine both values into a single number, the F1 score is often used. It is a kind of average of recall and precision, and only becomes high when both are high. An outlier on the low end pulls it down sharply. This prevents someone from shining with a one-sidedly optimized system.

Recall in products and headlines

The term appears everywhere systems are meant to filter something out. Search engines and the search function in company databases are measured by how many relevant documents they manage to display at all. Content moderation on platforms is likewise evaluated on recall: how much hate speech or how many fakes does the system detect, and what remains up? The figure also almost always appears in the requirements catalog for tenders for medical software.

When a company advertises a 99 percent detection rate, it’s worth asking a follow-up question. Sometimes recall is meant, sometimes precision, sometimes overall accuracy. These three figures can differ widely for the same system. Serious statements therefore always name at least two of them.

A note on the word itself: in English-language business news, 'recall' also refers to a recall action for defective products. This has nothing to do with the metric. In the AI context, it always refers to the hit rate, occasionally also called sensitivity.