Learn · Intermediate

Selective prediction and abstention: when an AI should decline to decide

Selective prediction is a way to let an AI system say “I should not decide this one” and send the case to a person, a stronger model, a search step, or a safer rule. It matters because many harmful AI failures come not from a model lacking any useful knowledge, but from a system forcing it to make a confident-looking choice outside the conditions where it has earned trust.

A standard classifier is built to answer every input. Give it an email and it must label it spam or not spam; give an agent a possible action and it must choose one. That setup silently assumes every case is equally answerable and every error costs the same. Real work is not like that. A false fraud block can annoy a customer; a missed fraud case can cost money. A mistaken medical triage decision can be much worse. An agent that is unsure whether it has permission to write to a production database should not turn uncertainty into a write.

Selective prediction changes the output space. Instead of only approve and reject, the system gets approve, reject, and abstain. The abstain option is not a shrug. It is an operational decision: ask for more evidence, call another tool, request human review, use a slower model, or decline the action. The system trades coverage—the percentage of cases it handles automatically—against risk on the cases it accepts.

An airport analogy helps. A routine bag can pass an automated lane. A bag with an unclear scan does not receive a random all-clear; it goes to secondary inspection. The screening machine need not be perfect to be useful. It needs to be good at separating easy cases from cases where a mistake is expensive. AI systems can use the same pattern: automate the high-confidence, low-risk majority and design a deliberate path for the remainder.

The research language for this is often selective classification or classification with a reject option. In Selective Classification for Deep Neural Networks, Yonatan Geifman and Ran El-Yaniv studied how a deep model can select a subset of inputs on which to predict while controlling risk. Their central measurement is a risk-coverage curve. At 100% coverage, the model answers everything and sees its ordinary error rate. As coverage decreases, it may reject uncertain examples and reduce error among the answers it keeps. There is no free lunch: higher reliability usually means more cases sent elsewhere.

Confidence is useful but treacherous. A model’s top probability is not automatically the probability it is correct. A system that says 90% confidence should be right roughly nine times out of ten among comparable predictions if it is calibrated. Guo, Pleiss, Sun and Weinberger’s calibration paper showed that modern neural networks can be badly overconfident and popularized simple post-hoc methods such as temperature scaling. Calibration is why a score like 0.82 can become an input to policy instead of decorative UI.

But calibration is not a substitute for abstention design. It can deteriorate when the input distribution shifts, when a prompt changes, when the available options are reordered, or when a model is used in a new domain. A local typed-decision model such as Jeff may emit a probability, but that probability is not permission to automate a consequential decision. Test it on representative cases, adverse cases, new wording, and cases that deserve an unknown outcome. Then choose a threshold according to the cost of an error and the capacity of the fallback queue.

For agents, the best abstention is often action-specific. An agent might be allowed to summarize a document at low confidence but required to abstain before sending email, spending money, changing access control, or deleting data. This is the bridge to agent identity and scoped credentials and sandboxing: uncertainty should narrow authority, not be hidden behind a fluent answer. A policy can also require evidence, such as a successful database read before a write or a human confirmation before a deployment.

A practical implementation has five pieces. First, define the action and its error cost. Second, measure confidence or another uncertainty signal on a held-out set that looks like production. Third, set a threshold and estimate expected coverage. Fourth, make the fallback concrete: human review, retrieval, a better model, or a harmless no-op. Fifth, monitor accepted-case error, abstention rate, and distribution shift after launch. A sudden drop in abstention may mean the model is becoming overconfident; a sudden increase may mean the input has changed or the threshold is too strict.

The honest limitation is that abstention cannot repair a system with no trustworthy fallback, and it can create delay or bias if hard cases are routed poorly. Nor does a model have direct access to metaphysical truth; it estimates uncertainty from patterns. Still, a well-designed reject option turns a brittle demand—“the AI must answer everything”—into a safer contract: automate what can be measured, expose what cannot, and give uncertainty somewhere useful to go.

Ground Truth. (2026, September 29). Selective prediction and abstention: when an AI should decline to decide. Ground Truth. https://groundtruth.day/learn/selective-prediction-and-abstention.html

@misc{groundtruth:selective-prediction-and-abstention,
  title  = {Selective prediction and abstention: when an AI should decline to decide},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/learn/selective-prediction-and-abstention.html}
}

Topics: evaluation · calibration · classification · agents · reliability