Is TP the same as accuracy?

No. Accuracy includes both correct positive and correct negative decisions.

Can a system have many TPs and poor precision?

Yes, if it also produces many false positives.

Does a true positive prove the model's explanation is correct?

No. A correct class decision can accompany an unsupported explanation.

AI Glossary / Data science and analytics / Advanced

True positive

Definition

A true positive is an outcome in which a classifier predicts the positive class and the reference label confirms that class. Usually abbreviated TP, it represents a correctly detected instance of the condition being tested. Positive identifies the designated target; it does not necessarily mean good, beneficial or welcome.

What it means

If an AI tool identifies articles that need an image caption, positive might mean caption missing. An article correctly flagged for a missing caption is a true positive. A correctly flagged problem is therefore a positive classification even though the underlying problem is undesirable.
A prediction becomes a recorded true positive through comparison with a trustworthy reference. A high model score is not enough. The score describes the model's output, whereas correctness requires evidence about the item.
The positive class must remain explicit. If the task instead identifies articles with complete captions, the same words refer to a different target and a different set of correct detections.

How it works

Define the target condition and decide what counts as one evaluated item. Establish reference labels using a suitable review or measurement procedure. Run the classifier, apply its decision rule and compare predicted classes with those labels.
Every item that is positive in both the prediction and reference contributes one true positive. Actual positives that receive negative predictions contribute false negatives. Predicted positives whose reference is negative contribute false positives.
Recall divides TP by TP plus false negatives. Precision divides TP by TP plus false positives. Recall asks how many target cases were found; precision asks how many positive predictions were correct.
Keep evaluation data separate from model development where possible. Repeatedly adjusting a model against the same test items can make the reported detections less informative about performance on new material.

Three practical examples

Website accessibility: An article-level checker correctly flags an article missing a required caption. The reference follows a defined caption policy, including legitimate exceptions.
Document organization: A classifier correctly identifies an invoice among incoming files. Positive means invoice, even if another document is equally important to the business.
Personal learning: A language-practice tool correctly identifies a sentence containing the particular grammatical error being tested. It need not detect every possible writing issue for that decision to be a true positive.

Think of it this way

Think of finding the right books on a reading list. Every correctly selected required book is a true positive. That count does not say whether you overlooked other required books or added unrelated ones. Those two questions correspond to missed targets and false alarms, which need their own counts.

Worked example

An evaluation set contains 120 articles. Reviewers identify 30 that are missing required captions and 90 that are not. A checker flags 36 articles. Of those, 24 genuinely need captions and 12 do not. It misses the remaining 6 articles with missing captions.
Thus TP = 24, FP = 12, FN = 6 and TN = 78. The counts sum to 120.
Recall is 24 / 30 = 80%. Precision is 24 / 36, approximately 66.7%. The system finds four-fifths of the target articles, but one-third of its alerts are unnecessary under the reference policy.
Suppose a lower threshold flags ten additional articles, including four genuine targets and six articles that do not need captions. TP rises to 28, FP to 18, FN falls to 2 and TN falls to 72. Recall becomes approximately 93.3%, while precision becomes approximately 60.9%.
The additional four true positives cost ten extra reviews. Whether that trade-off is worthwhile depends on the workflow. The arithmetic does not select a universally best threshold.

Illustrative case study

Fictional education business Linden Courses pilots a caption checker. Its dashboard initially shows only the number of issues found. The team notices that longer articles with more images contribute disproportionately to that number.
Editors decide to evaluate at article level: an article is positive if it has at least one missing required caption. They separately retain the number of missing captions for estimating repair effort. One article with five omissions now contributes one true positive to article classification, rather than five.
Linden tests the revised reporting on independently reviewed material and inspects both missed articles and unnecessary alerts. This makes the success count interpretable and prevents a change in counting units from masquerading as a model improvement. The business and workflow are illustrative, not a published performance claim.

Common misconceptions and mistakes

A positive prediction is not automatically a true positive. Verification can reveal a false alarm.
More true positives do not necessarily mean a better model. A larger dataset, more common target condition or lower threshold can increase the count.
Recall and precision are not alternative names for TP. They are ratios using different comparison groups.
Correct detection does not prove that an intervention worked. Identifying a missing caption and successfully repairing it are separate outcomes.
Counts require a clear unit. Mixing article-level decisions with individual caption detections produces a misleading confusion matrix.

Business use

True positives help estimate useful detections and the workload associated with acting on them. Pair the count with missed targets, false alerts, evaluation size and target prevalence. Track downstream outcomes separately, such as verified repairs or time spent reviewing. A model can correctly identify many problems while the organization lacks capacity to resolve them, so classification success alone is an incomplete measure of business value.

Check your understanding

Why is recall 24 / 30 in the example?

Answer: Thirty articles actually need captions.

Why is precision 24 / 36?

Answer: Thirty-six articles receive positive predictions.

How many article-level TPs does one correctly flagged article with five omissions contribute?

Answer: One, under the stated unit.

Related terms

Learn next

Compare two decision thresholds using a complete confusion matrix. Translate the additional correct detections and false alerts into actual review work before choosing an operating point.

Confusion matrix

Sources

Google, Machine Learning Crash Course, Thresholds and the Confusion Matrix: positive-class decisions and evaluation outcomes. https://developers.google.com/machine-learning/crash-course/classification/thresholding Google, Classification: Accuracy, Recall, Precision, and Related Metrics: definitions of recall and precision and threshold trade-offs. https://developers.google.com/machine-learning/crash-course/classification/accuracy-precision-recall The caption examples, numerical comparison and Linden Courses case are original illustrations. Their figures are not performance claims about a commercial system.

Previous: True online TD(lambda) Next: True positive rate