TruthfulQA — AI Glossary

Reference details

Related terms

• Calibration
• Hallucination
• HellaSwag

How it works

1. Questions are constructed adversarially around common human misconceptions, so the plausible-sounding answer is the wrong one and imitating confident text is penalised rather than rewarded.
2. That construction is the point: a model trained to predict likely text will reproduce the common false belief, so the benchmark measures resistance to a specific failure of the training objective.
3. Two properties are scored together. Truthfulness asks whether the answer avoids asserting something false; informativeness asks whether it says anything at all.
4. Scoring both is necessary because they are trivially separable — a model that declines every question is perfectly truthful and useless, and the pair is what makes the metric meaningful.
5. Both generative and multiple-choice formats exist, and they are not interchangeable: the choice format floors at chance and the generative format requires a judgement about what the answer asserted.
6. The question set is small and public, so contamination is the default assumption and the per-question intervals are wide.

Example: Why truthfulness alone is a metric you can win by refusing

Take the 817 questions. A model answering 'I don't know' to every one asserts nothing false, so it scores 100% on truthfulness and 0% on informativeness — which is why the pair must be reported together and why a single truthfulness figure is not a result. Now the precision available. At 817 questions and a score near 50%, the standard error of a proportion is sqrt(0.5 x 0.5 / 817) = sqrt(0.000306) = 0.0175, about 1.7 points, and the standard error on a difference between two models is sqrt(2) x 1.7 = 2.5 points. So gaps under about 2.5 points are indistinguishable from sampling, and a reported 1-point improvement on this set carries no information. Resolving 1 point would need the difference's standard error near 0.5, which requires sqrt(2 x 0.25 / n) = 0.005, so n = 0.5 / 0.000025 = 20,000 questions — roughly 24 times the set. Combined with the contamination risk on a small public set, the honest use is as a qualitative check on a known failure mode rather than as a figure to rank models by.

Commonly confused with

This measures resistance to common misconceptions, not factual accuracy in general and not hallucination. Factual accuracy would sample knowledge broadly; this samples adversarially where the likely answer is false, so it is a targeted probe rather than a survey. Hallucination is a different failure — fabricating specifics that were never in the training data or the context — while the failure here is reproducing a false belief that genuinely is common in text. A model can be strong here and still fabricate citations, which is why the two need separate measurement.

When to use it

Reach for it when:

• As a targeted check on misconception reproduction, particularly in health, legal and financial domains where the benchmark concentrates.
• With truthfulness and informativeness reported together, since either alone is gameable.
• Qualitatively, by reading which misconception clusters a model fails, which is more useful than the aggregate.
• Alongside separate hallucination and factual-accuracy measurements, which it does not substitute for.

Reach for something else when:

• Ranking models on small differences, where 2.5 points is one standard error on 817 questions.
• As a general factual-accuracy measure, since the questions are adversarially selected rather than representative.
• Reporting truthfulness without informativeness, which a model wins by declining to answer.
• Without accounting for contamination, which on a small public set is the default assumption.

Primary source

• Lin et al., TruthfulQA: Measuring How Models Mimic Human Falsehoods (arXiv)

Continue your research

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

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