Abstaining when unsure lowers wrong commitments and boosts accuracy
A simple trick: let AI abstain when it lacks solid evidence, and only answer when it’s well supported; results vary with gate strictness.
- Publication
- arXiv
- Stage
- Preprint
- What we read
- Summary of the paper
- Authors
- Ali Şenol
- Universities and research institutions
- Not yet supplied in verified metadata; the Brief does not guess.
What they did and found
We tested three AI styles that can refuse to answer when evidence is weak. At a 0.90 confidence cut-off, all three reduced wrong commitments and improved accuracy across eleven models on an 817-question test, with some refusals.
Why it matters
In use, an AI that can say 'I don’t know' is safer to deploy, especially where a wrong answer costs more than asking a human.
Two ideas guided the work: first, the AI should show what it needs and how sure it is; second, a gate decides to answer or abstain.
A company could start with a conservative gate, abstain more, and adjust the threshold to balance speed, coverage, and risk for important tasks.
Original sources · 1
Check the original paper for its authors, methods, version and access terms.
What remains uncertain
The study used one test type; results may not apply to other areas, and confidence signals need better calibration.
What is your take?
Ask a question, add useful context or share a different perspective. Keep the conversation respectful and grounded.