Selective Prediction Explainer
Selective prediction allows an AI model to produce an answer for some inputs and abstain on others. A selection function, often based on confidence or uncertainty, decides which cases the system will handle automatically.
As the system becomes more selective, error on answered cases may decrease while coverage also decreases. The resulting risk-coverage curve shows the tradeoff between how many cases receive predictions and how reliable those predictions are.
Abstained cases need a destination such as human review, a safer rule, a second model, or a request for more information. Deferral can create delays and unequal burdens if difficult cases cluster within particular groups.
The selection score must be validated on realistic data. A model that is confidently wrong will not defer the right cases, and a threshold tuned on one environment may fail after shift or changes in downstream costs.
Selective prediction lets a model answer only some cases and defer or abstain on others when its selection rule indicates greater uncertainty or risk.
Key terms include selective prediction, selection function, abstention, deferral, confidence, risk, coverage, threshold and review.
Answering fewer cases can reduce error among the cases the model accepts.
The organization must decide how much deferral its workflow can safely support.
Abstention is useful only when a safe and timely fallback receives the case.
Teams should measure who is deferred and whether the fallback performs well.
Explore more "Explainers"
Discover additional explainers across politics, science, business, technology, and other fields. Each explainer breaks down a complex idea into clear, everyday language—helping you better understand how major concepts, systems, and debates shape the world around us.