Abstention is the deliberate decision by a machine learning model to withhold a prediction for a specific input, signaling that the model's confidence is below a reliability threshold and the input should be handled by an alternative mechanism—typically human review, a more specialized model, or a conservative default action. Abstention is the operational implementation of selective prediction, converting uncertainty awareness into actionable "I don't know" decisions.

Why Abstention Matters in AI/ML: Abstention provides the critical safety mechanism that prevents unreliable AI predictions from being acted upon in high-stakes applications, acknowledging that an honest "I don't know" is far more valuable than a confident wrong answer.

• Confidence-based abstention — The simplest form: abstain when max softmax probability < threshold τ; setting τ = 0.95 means the model only predicts when at least 95% confident; the threshold is tuned to achieve the desired accuracy-coverage tradeoff on validation data • Uncertainty-based abstention — More sophisticated: abstain based on epistemic uncertainty (ensemble disagreement, MC Dropout variance) rather than raw confidence; this catches inputs where the model is uncertain even if individual predictions appear confident • Cost-sensitive abstention — Different errors have different costs (e.g., false negative cancer diagnosis vs. false positive); abstention thresholds are set per-class based on the relative cost of errors versus the cost of human review • Learned abstention — A dedicated abstention head is trained jointly with the classifier, learning directly when to abstain rather than relying on post-hoc thresholding; this can capture subtle patterns of model unreliability invisible to simple confidence scores • Cascading systems — Abstention triggers escalation through a cascade: fast cheap model → slower accurate model → human expert; each stage handles cases within its competence and abstains on harder ones, optimizing cost-accuracy across the system

Abstention is the essential safety valve for AI systems, transforming uncertainty quantification into actionable decisions that prevent unreliable predictions from reaching end users, enabling honest, trustworthy AI deployment where the system's silence on uncertain cases is as informative and valuable as its predictions on confident ones.