Abstention — AI Glossary

Reference details

Related terms

• Refusal
• Calibration
• Hallucination

How it works

1. The model must have a signal correlated with its own reliability — low likelihood, disagreement across samples, or absent retrieved evidence.
2. A threshold on that signal decides between answering and declining, which makes abstention a tunable operating point rather than a behaviour.
3. Grounded systems get the cleanest signal: if retrieval returned nothing relevant, there is a concrete basis for declining.
4. The prompt must make declining an allowed response, because a model asked to answer will answer rather than volunteer uncertainty.
5. Threshold choice trades two errors against each other — declining answerable questions, and answering unanswerable ones — and the right point depends on which costs more.

Example: The threshold is a cost decision, not a quality one

Of 1,000 questions, 100 are unanswerable from the available evidence. A conservative threshold might decline 150: it catches 95 of the 100 and also declines 55 answerable ones. A permissive threshold declines 40: it catches 38 and wrongly declines only 2, leaving 62 confabulated answers. Neither is better in the abstract — in a medical or legal setting the 62 wrong answers are unacceptable, and in a search assistant 55 unnecessary refusals would make the product useless.

Commonly confused with

Routinely conflated with refusal, and they come from different places. Refusal is policy-driven: the model will not answer because the request is out of bounds, and the correct response is the same regardless of how much it knows. Abstention is epistemic: the model would answer if it had reliable grounds and does not, so the same question becomes answerable once evidence is supplied. Reporting them as one 'did not answer' rate hides which lever applies — a policy change or better retrieval.

When to use it

Reach for it when:

• A wrong answer costs more than no answer, as in clinical, legal or financial contexts.
• Grounded generation, where absent evidence gives a concrete basis for declining.
• Calibration is being measured, since abstention is where it becomes observable.

Reach for something else when:

• Describing policy-driven non-answers, which are refusals.
• Where an unnecessary decline is worse than a hedged answer, as in exploratory search.
• Without a reliability signal, where declining becomes arbitrary rather than calibrated.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Abstention — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path