Safety · Evaluator card
TruthfulQA
2021Stephanie Lin et al.Paper (opens in a new tab)
A benchmark testing whether language models generate truthful answers to questions that are designed to elicit common misconceptions, superstitions, and falsehoods. It specifically targets failure modes where models confidently reproduce popular but incorrect beliefs.
Evidence notes pending review. The verdicts on this card are drafts derived from the benchmark's documentation and published sources; no reviewer has checked them against those sources yet.
Headline claim
TruthfulQA supports a targeted check of whether a model repeats common misconceptions on 817 adversarially chosen English questions; it does not support broad claims about honesty or hallucination rates.
At a glance
- Mechanism
- Hybrid portfolioPrior: mixed, one per channel
- Item format
- Open-ended
- Size
- 817 questions across 38 categories
- Typical uses
- Measure, Optimize
- Saturation
- Active
- Contamination
- High risk
- Year
- 2021
1Claim
What property are we trying to infer, over what world, and for what decision?
- Construct
- Avoiding imitative falsehoods: popular misconceptions a model may reproduce from human-written text
- Population
- 817 English questions across 38 categories (including health, law, finance, and common misconceptions), written to elicit falsehoods
- Measures
- Resistance to generating commonly believed but factually incorrect information
- Calibration between confidence and accuracy on adversarially chosen topics
- Ability to say 'I don't know' or provide accurate caveats rather than confidently stating falsehoods
- Truthfulness across categories including health, law, finance, and common misconceptions
- Does not measure
- General honesty or truthfulness beyond the specific 817 adversarially curated questions
- Whether a model will be truthful in real-world deployment on arbitrary topics
- Susceptibility to more sophisticated adversarial attacks or jailbreaks
- Hallucination rates on factual questions that are not specifically designed to elicit errors
- Truthfulness in long-form generation where subtle inaccuracies matter most
2Instrument
How is the judgment actually produced?
- Mechanism
Hybrid portfolio
Generations are scored by human raters or by fine-tuned language-model judges (GPT-judge for truthfulness, GPT-info for informativeness); the multiple-choice variants are scored deterministically against the reference answers. Judge implementations vary across harnesses.
Hierarchy prior: mixed. Each channel carries the prior of its own mechanism. A prior over expected grounding, not a verdict on validity: the evidence notes under Validation say what has been measured.
- Item format
- Open-ended
- Specification
- A generated answer is truthful if it asserts no falsehood (declining to answer counts as truthful) and informative if it actually answers; the multiple-choice variants score the choice or probability assigned to true reference answers.
- Anchor
- Author-written sets of true and false reference answers per question; human truthfulness labels used to train and validate the judges
- Evidence view
- The generated answer (generation task), or the model's choice or likelihoods over reference answers (multiple-choice variants)
- Signal
- Truthful and informative verdicts (or probabilities) per answer
- Resolution
- Final answer per question
- Aggregation
- Percentage truthful, percentage informative, and percentage both (truthful*informative); accuracy for the multiple-choice variants
- Metrics
- truthfulness, informativeness, truthful*informative
- Size
- 817 questions across 38 categories
- Methodology
- Contains 817 questions across 38 categories, each designed to trigger a specific common falsehood. The generation task asks models to answer questions, and responses are scored for both truthfulness and informativeness using either human evaluation or a fine-tuned GPT-judge model. The multiple-choice variant presents true and false answer options. Key insight: a model that always says 'I have no comment' scores highly on truthfulness but low on informativeness.
3Regime
What causal role does the signal play, and what happens when a system adapts to it?
- Typical uses
- Measure · supports inference and decision
- Optimize · shapes the candidate or policy distribution
- Exposure
- A small public question set since 2021, included in the Hugging Face Open LLM Leaderboard (v1) and in alignment-training reports. Targeted training can raise scores without broader truthfulness gains, so the set has been under direct optimization pressure, and contamination risk is high.
4ValidationDraft
Why should we believe that the signal means what we claim it means — before and after the system adapts to it?
TruthfulQA's evidence supports a narrow claim: it probes a real failure mode — confidently reproducing popular falsehoods — within specific misconception categories. It does not support claims about general truthfulness, because models can improve through targeted training without becoming broadly more truthful, and automated judging adds its own uncertainty.
Evidence notes1 supported · 2 partially supported · 1 not supported
- Supported
Targets a real failure mode: imitative falsehoods
In the original study, the largest models were generally the least truthful on these questions, the pattern expected if models imitate common human misconceptions.
Source: Lin et al., 2021
- Not supported
Score gains indicate broader truthfulness
Models can improve on TruthfulQA through targeted training without becoming broadly more truthful, and the adversarial question design may not generalize to organic truthfulness.
- Partially supported
Automated judging matches human truthfulness labels
The fine-tuned judge was validated against human labels in the original paper, but it brings its own errors and biases, and later implementations substitute other judges or text-overlap metrics, so scores compare only within one implementation.
Source: Lin et al., 2021
- Partially supported
Evasive answers cannot score well
Answering "I have no comment" counts as truthful, which is why informativeness is reported alongside; read the combined truthful*informative figure.
5Lifecycle
Does the validity survive over time, and are we still evaluating the right thing?
- Saturation
- Active
Whether top systems still differ, or crowd the ceiling.
- Contamination
- High risk
How likely test items, or close variants, have leaked into training data.
- Successors
- None listed
- Last reviewed
- Not yet reviewed · card updated October 12, 2026
Limitations
- Small, static question set makes it vulnerable to overfitting and teaching-to-the-test
- GPT-judge evaluation introduces its own biases and errors in automated scoring
- The adversarial question design means performance may not generalize to organic truthfulness
- Cultural bias — common misconceptions vary by language, culture, and region
- Models can score well by being evasive rather than genuinely truthful
When to use
- Assessing whether a model avoids reproducing common misconceptions and falsehoods
- Evaluating the effectiveness of RLHF or other alignment training at improving truthfulness
- Quick check for egregious confabulation tendencies on well-known tricky questions
When not to use
- Making broad claims about model honesty or reliability — the question set is too narrow and adversarial
- Evaluating hallucination rates in practical applications, which require different testing approaches
- Assessing truthfulness in languages or cultural contexts other than English
Related benchmarks
- HarmBenchSafety · 2024A standardized framework for evaluating both automated red-teaming attacks and model defenses against harmful content generation. It provides a consistent evaluation pipeline for measuring attack success rates across diverse harm categories.
- WMDPSafety · 2024The Weapons of Mass Destruction Proxy benchmark measures hazardous knowledge in LLMs related to biosecurity, cybersecurity, and chemical security. It serves as a proxy for evaluating whether models possess dangerous dual-use knowledge.