Of all the times a model predicted positive, the fraction it was actually right — a measure of how trustworthy positive predictions are.
Precision answers the question: when the model said "yes," how often was it correct? Formally, it is the number of true positives divided by the total number of positive predictions (true positives plus false positives). A precision of 0.90 means that 90% of the model's positive predictions were correct, and 10% were false alarms.
Precision is the metric to prioritise when false positives are costly. A link spam detector with low precision flags legitimate links as spam — causing unnecessary disavow work and potentially hurting SEO. A high-precision detector may miss some spam (low recall) but everything it does flag is genuinely problematic. The cost of each error type determines which metric to optimise.
For DEJAN's AI content detection model, high precision means that when it flags content as AI-generated, the flag is reliable — clients can act on it without manually reviewing every prediction.
Precision and recall trade off against each other. Raising the decision threshold makes the model more conservative — it predicts positive less often, so when it does, it's more likely to be right (higher precision) but it misses more actual positives (lower recall). The F1 score combines both into a single number. DEJAN's training pipelines track all three and select checkpoints by F1, with precision and recall reported alongside for transparency.
