Recall

The Precision article closed with a specific gap it couldn’t fill on its own: it says nothing about how many real positives a model actually caught. That’s exactly the question recall answers.

The simple definition

Recall measures what fraction of all the genuinely positive cases a model actually managed to catch. If there are 100 real spam emails in a dataset, and a spam filter correctly flags 80 of them (missing 20), its recall is 80/100 = 80%. Recall doesn’t care at all about legitimate emails wrongly flagged as spam — it’s entirely focused on the opposite question from precision: of everything that was genuinely positive, how much did the model actually find?

Using the confusion matrix’s four cells, recall’s formula looks like this:

recall = True Positives / (True Positives + False Negatives)

In plain terms: of everything that was genuinely positive (True Positives plus False Negatives — the real positives the model missed), what fraction did the model actually catch?

Precision and recall, side by side

These two metrics are close cousins, both built from the same confusion matrix, but they ask genuinely opposite questions, and it’s worth holding them side by side clearly:

• Precision asks: of what the model flagged, how much was correct? (Punishes False Positives.)
• Recall asks: of what actually existed, how much did the model find? (Punishes False Negatives.)

flowchart LR
    A[100 real spam emails exist] --> B[Model catches 80: True Positives]
    A --> C[Model misses 20: False Negatives]
    B --> D[Recall = 80 / 100 = 80%]
    C --> D

ANALOGY vs. TECHNICAL REALITY

Analogy: Think of a fishing net designed to catch a specific rare fish species out of a lake. High recall means the net catches nearly every one of that species actually present in the lake — few slip through unnoticed. A net with low recall lets most of the target fish swim right through, even if everything it does catch turns out to genuinely be the right species (which would be high precision, but poor recall).

Where this breaks down: A fishing net’s “catching” is a single physical mechanism. Recall, like precision, is a purely calculated, after-the-fact statistic — comparing a model’s predictions against the full, known set of real positives, which requires reliable Ground Truth to even measure in the first place, unlike a fisherman who can directly see what swam past uncaught.

Where high recall matters most

Recall becomes the priority whenever missing a real positive case is the costly, dangerous mistake — more costly than the inconvenience of an occasional False Positive. Cancer screening is the textbook example: missing a real cancer case (a False Negative) can be life-threatening, while an unnecessary follow-up test for a healthy patient (a False Positive) is a real but far smaller cost. A security system scanning for serious threats generally prioritizes recall for the same reason — better to investigate a few false alarms than to let a genuine threat slip through entirely undetected.

Calculate recall step by step

In the shared screening example, 20 people truly have the condition:

18 were correctly found (TP)
 2 were missed         (FN)

recall = TP ÷ (TP + FN)
       = 18 ÷ (18 + 2)
       = 18 ÷ 20
       = 0.90
       = 90%

Plain-English reading:

This model finds 90 of every 100 people who truly have the condition.

Recall does not use true negatives or false positives in its formula. Its attention stays on the actual positive cases.

Recall is also called sensitivity or the true positive rate. The name changes across fields, but the calculation is the same.

Improve recall by reducing false negatives

Lowering the decision threshold usually marks more cases positive. This may catch more true cases and improve recall, but it can also create more false positives and reduce precision. The best threshold depends on the cost of each mistake.

A concrete example, layered

For a simple beginner example: a smoke detector needs very high recall — missing an actual fire (a False Negative) is a potentially catastrophic failure, while an occasional false alarm from burnt toast (a False Positive) is merely annoying, not dangerous, so smoke detectors are deliberately tuned to err toward triggering rather than staying silent. For a production example: a medical diagnostic model screening for a serious disease is typically tuned to maximize recall even at some cost to precision, since a missed diagnosis (False Negative) can have severe consequences for a patient, while a False Positive simply leads to a follow-up test that ultimately clears the patient — a real, deliberate trade-off healthcare AI teams make explicitly when tuning these models, echoing the medical-versus-spam cost asymmetry first raised in the Evaluation article.

Why recall alone isn’t the whole picture either

This mirrors precision’s own limitation from the opposite direction, and it’s worth stating plainly. A model can achieve perfect recall trivially — simply by flagging everything as positive. If a spam filter marks every single email as spam, it will have caught 100% of the real spam (perfect recall) while also wrongly flagging every legitimate email too (terrible precision). Recall alone doesn’t punish this kind of over-eager, indiscriminate flagging — which is exactly why neither precision nor recall should be used entirely alone, and why the next article, F1 Score, exists to combine them into one balanced measure.

Real AI-agent example: safety favored recall

The Operator System Card reports 99% recall for its prompt-injection monitor on 77 red-team attacks, with one borderline attempt missed.

Why favor recall here? A missed prompt injection could influence an AI agent’s actions. The team therefore wanted to catch nearly every attack, even though stronger detection can sometimes interrupt a benign task.

This is a real precision–recall decision:

Higher recall → fewer dangerous attacks missed
Possible cost  → more safe pages may be flagged

The “right” balance follows the consequences of errors. A photo-search tool and an agent capable of taking actions may reasonably use different thresholds.

Common misconception

A frequent beginner assumption: that maximizing recall is always the safer, more responsible choice, since missing something sounds worse than a false alarm. As the “flag everything” example above shows, this isn’t automatically true — a model tuned purely for maximum recall, with no regard for precision, becomes practically useless the moment it starts flagging nearly everything as positive. The right balance between precision and recall genuinely depends on the specific task and the specific real-world cost of each type of mistake — there’s no universally “safer” direction to lean.

Where this fits in what comes next

You now understand both halves of the precision-recall trade-off: how trustworthy a model’s positive flags are, and how thoroughly it catches real positives. The final article in this phase, F1 Score, covers the standard way of combining both into a single, balanced number — useful whenever you need one score that fairly reflects both concerns at once.

In one sentence

Recall measures how thoroughly a model catches the real positive cases that actually exist, and it’s the metric to prioritize whenever missing a case is the costlier mistake — though, exactly like precision from the opposite direction, it can be trivially gamed and shouldn’t be trusted as the sole measure of a model’s quality.

Related Terms

• Precision
• Confusion Matrix
• Evaluation
• F1 Score
• Accuracy