---
title: "What is abstention in AI agents?"
description: "Abstention is a system declining to answer when evidence is missing rather than guessing. Why AI agents need it and how the signal works."
canonical: https://past.dev/glossary/abstention
last-updated: 2026-09-02
---
# Abstention

Source: https://past.dev/glossary/abstention

Abstention is a system declining to answer because the stored evidence does not establish an answer, instead of guessing. For AI agents abstention is a capability. A memory system that always returns its best match will present weak matches as answers, and the agent will act on them. A system that can say the evidence is absent lets the agent ask, escalate, or search elsewhere instead.

## Why guessing is worse

Retrieval always returns something: the ranking has a top result whether or not the answer exists in storage. Without abstention, that top result gets treated as the answer. The failure compounds in three ways:

1. A guess carries the same surface confidence as a supported answer, so downstream steps cannot tell them apart.
2. Agents act. A guessed refund amount becomes an issued refund; a guessed commitment becomes a promise to a customer.
3. Agents write back to memory. A guess stored as a fact resurfaces in later recalls as if it were established. See [memory poisoning](/glossary/memory-poisoning).

The asymmetry drives the design. A missing answer costs a follow-up question. A wrong answer that reaches a customer, a ledger, or a stored record costs the correction, the lost trust, and the cleanup of everything derived from it in the meantime.

## How abstention is signaled

An empty result list is a poor signal, because it is ambiguous: the fact may be absent from storage, the query may have missed it, or retrieval may have failed partway. Abstention needs an explicit [evidence status](/glossary/evidence-status) on the response, distinct from both a supported answer and an infrastructure failure. In past.dev, the `NoKnownSupport` status value is the abstention signal: recall states that the stored evidence does not establish an answer. A well-designed status set also reports retrieval failure separately, so a failed lookup is never mistaken for confirmed absence.

An abstention is actionable. Typical agent responses, in order of preference:

1. Ask the user for the missing information, stating what was searched.
2. Reformulate the query: widen the time range, try the entity's other names, drop a filter.
3. Consult a different source, such as a live system or a human escalation path.
4. Record that the question was unanswerable at this time, so the gap itself becomes known.

## Abstention in evaluation

Benchmark suites include questions whose answers are absent from the stored input on purpose. A system that answers every question is guessing on those. Abstention is therefore evaluated as an ability of its own: declining the unanswerable questions while still answering the supported ones. A decline on an answerable question is also a failure, so the evaluation separates the two mistakes rather than collapsing them into one number. The [benchmark methodology](/benchmarks/methodology) describes how these cases are constructed and scored.

## Related concepts

- **[Evidence status](/glossary/evidence-status)**: the response label that carries the abstention signal.
- **[Grounding](/glossary/grounding)**: tying answers to evidence, which abstention protects when evidence is missing.
- **[Recall precision](/glossary/recall-precision)**: returning less irrelevant material, the retrieval-side counterpart.
- **[Memory poisoning](/glossary/memory-poisoning)**: what guesses written back to memory turn into.

## Frequently asked questions

### What is abstention in AI?

Abstention is a system choosing to say it does not know because its evidence does not establish an answer. It replaces a guess with an explicit signal the caller can act on.

### Why is abstention better than a best guess?

A guess looks identical to a supported answer, so downstream steps treat it as true. An abstention tells the agent to ask the user, escalate, or look elsewhere before acting.

### How do benchmarks test abstention?

Test sets include questions that cannot be answered from the stored input. Systems are checked on declining those questions while still answering the ones the evidence supports.

## Related

- [Evidence status](https://past.dev/glossary/evidence-status)
- [Grounding](https://past.dev/glossary/grounding)
- [Recall precision](https://past.dev/glossary/recall-precision)
- [Contradictory facts](https://past.dev/guides/contradictory-facts)
- [Memory API quickstart](https://past.dev/docs/memory-api/quickstart)