True Positive Rate

An AI model makes decisions: sick or healthy, spam or not spam, fraud or no fraud. Sometimes it’s right, sometimes it’s wrong. The true positive rate measures a specific error: how many of the truly positive cases — that is, those the model is actually supposed to detect — are in fact detected. For example, if a model makes cancer diagnoses and 100 patients are genuinely sick, but the model classifies only 80 of them as sick, the true positive rate is 80 percent. The remaining 20 percent were missed — these are called false negatives. So the true positive rate tells you: how thoroughly does the model work at finding what it’s supposed to find?

What’s at stake when the rate is too low

A low true positive rate means the model simply lets many relevant cases slip through. In medicine, this can be life-threatening: a diagnostic system that regularly misses real cancer cases is barely better than no system at all. The same applies to security software that fails to reliably detect malware.

It’s important to note: a high true positive rate alone doesn’t make a model good. A model that simply classifies everything as positive would have a true positive rate of 100 percent — but it would simultaneously label every healthy person as sick. That’s why the true positive rate is always considered together with other metrics, especially the false positive rate, meaning the proportion of cases incorrectly classified as positive.

Calculation and the relationship with the threshold value

The formula is simple: you divide the number of correctly identified positive cases by the total number of actually positive cases. The result always lies between zero and one, but is often expressed as a percentage. Mathematically: TPR = true positives / (true positives + false negatives).

An AI model usually doesn’t output a direct yes/no value for each case, but rather a probability — for example, 0.73 for a “73% chance of fraud.” Only once a certain threshold is reached does the system classify the case as positive. Lowering this threshold increases the true positive rate: the model becomes more generous and detects more genuine cases. At the same time, however, the number of false alarms also increases. Raising the threshold has the opposite effect. This trade-off can be visualized in what is known as an ROC curve — a graphical representation.

In practice, developers choose the threshold depending on the field of application. In cancer detection, missing a real case is worse than triggering a false alarm — so a low threshold is chosen, accepting a high true positive rate. For an automated loan application, the trade-off might be weighed differently.

True positive rate in products and debates

The term comes up wherever AI systems are meant to detect or classify something. Facial recognition used by police is measured by how reliably it identifies wanted persons. Spam filters are supposed to catch as many unwanted emails as possible. Medical image analysis is supposed to not miss tumors.

In political and societal debates about AI fairness, the true positive rate plays a particular role. Studies have shown that some facial recognition systems achieve a significantly lower true positive rate for people with darker skin than for lighter-skinned people — the model recognizes them correctly less often. This is a concrete example of how a technical metric becomes a political issue.

In financial news, the term frequently appears in connection with fraud detection systems at banks or AI-supported diagnostic assistance in medical technology. When a company claims its model detects 95 percent of all fraud cases, it means precisely the true positive rate — and often fails to mention how many false alarms it produces at the same time.