TruthfulQA and Measuring Honest Answers
9 min read · updated August 4, 2026
TruthfulQA is around 800 questions written so that a person repeating a common misconception would get them wrong. It measures one specific failure — reproducing a falsehood that is popular in the training data — and a model can score perfectly on it by refusing to answer anything, which is why truthfulness is never reported alone.
What TruthfulQA is
TruthfulQA was published by Lin, Hilton and Evans in 2021. The set is roughly 800 questions across several dozen categories — health, law, finance, politics, conspiracies, misconceptions, fiction mistaken for fact, and others — each with a set of reference answers marked true and a set marked false.
The questions were written adversarially. The authors composed questions they believed a language model would answer falsely because a falsehood was common in human writing, then filtered: questions a model of the day answered correctly were largely dropped, and questions it answered falsely were kept. A subset was then validated by independent human annotators.
That filtering step is the benchmark’s defining property, and most of what follows is a consequence of it.
Imitative falsehood is a narrow target
The paper’s term for what it measures is imitative falsehood: a false statement the model produces because the training distribution contains it, not because the model lacks the information. The distinction matters because it separates TruthfulQA from every other correctness benchmark.
| Failure | Description |
|---|---|
| Imitative falsehood | The model repeats a popular misconception. More capable models can get worse at this, because they model the training distribution more faithfully — this is the paper's headline claim and the reason the benchmark is interesting. |
| Knowledge gap | The model does not know. Measured by MMLU-style factual tests, not by this one. |
| Confabulation | The model invents a specific, plausible, non-existent fact — a citation, a statute, a version number. A different failure with a different fix; see the hallucination taxonomy. |
| Reasoning error | The model derives a wrong answer from correct premises. Measured by maths and logic sets. |
Only the first is TruthfulQA’s subject. A model that scores well here can still fabricate references freely — the taxonomy in types of hallucination separates the six things people call by one name, and fabricated citations is a different measurement problem entirely.
Three scoring modes, three different numbers
TruthfulQA can be run three ways, and papers frequently report one without saying which. They are not comparable.
| Mode | Description |
|---|---|
| Generation | The model answers freely and a grader decides whether the answer is truthful and whether it is informative. The original grader was a model fine-tuned on the authors' human judgements — a dependency that has aged badly, since reproducing it exactly is now difficult. Most current runs substitute a different judge, which changes the numbers. |
| MC1 | Multiple choice with one true reference answer among several false ones. Scored by whether the model assigns the highest likelihood to the true option. A single accuracy. |
| MC2 | Multi-true multiple choice: several options are true. The score is the total normalised probability mass the model puts on the true set. Not an accuracy — a probability mass, which is why MC2 values sit in a different range from MC1 and cannot be compared with them. |
MC2, worked. Six options; the first three are true. Model's normalised probabilities over the six options: true: 0.31, 0.14, 0.05 sum = 0.50 false: 0.30, 0.15, 0.05 sum = 0.50 MC2 score = 0.50 MC1 on the same item asks only: is the single designated true answer the argmax? Here the argmax is 0.31, which is a true option, so MC1 = 1. One item. Two modes. 1.00 and 0.50.
The generation mode is the one the benchmark was designed around and the only one that tests what a user would experience. The multiple choice modes exist because generation grading is expensive and judge-dependent, and they measure a proxy: whether the model ranks truth above falsehood among options it was handed, which is a weaker claim than producing a true answer unprompted.
Why one number is always a lie here
This is the design feature worth taking away from TruthfulQA even if you never run it. The benchmark scores two axes, and the trivial way to maximise one destroys the other.
Strategy: answer "I have no comment." to all 800 questions. Truthfulness: 100% (no false statement was made) Informativeness: 0% (no question was answered) Strategy: answer every question confidently from popular belief. Truthfulness: low Informativeness: 100% Neither is a good model. The benchmark reports both, and the pair — or the product, or the truthful-and-informative rate — is the score.
A published TruthfulQA number that gives only truthfulness is missing the denominator that makes it meaningful. The right figure to quote is the rate of answers that are both truthful and informative.
This generalises well beyond this benchmark. Any safety or honesty metric with a trivial degenerate maximiser must be reported alongside the metric that degenerate strategy destroys — the same structure appears in safety and refusal benchmarks, where a model that refuses everything scores perfectly on harm and uselessly on over-refusal.
The standing critique of the question set
TruthfulQA is widely used and widely criticised, and the criticisms are substantive rather than sour.
- The set was filtered against one model in 2021. Questions were kept because a particular model of that era answered them falsely. That makes the set a portrait of one model’s failure modes as much as of a general phenomenon, and it means later models are being tested on an adversarial set built against an ancestor. Whether the questions generalise is an empirical question the construction cannot settle. The same objection applies to BBH’s selection rule.
- Some gold answers are contestable. A set spanning law, health, economics and politics contains questions where the authoritative answer depends on jurisdiction, on date, or on framing. Reference answers fixed in 2021 encode a particular resolution of those, and a model giving a defensible different answer is graded wrong.
- Reference-based grading penalises better answers. An answer that correctly explains why the question rests on a false premise may match no reference. This is the general weakness of reference-based evaluation and it is why the generation mode needs a judge rather than string matching.
- It is thoroughly contaminated by now. The dataset, its reference answers and years of analysis of it are public. A model trained after 2021 has almost certainly seen the questions with their correct answers attached, which is precisely the exposure that makes a score uninterpretable — see contamination.
What it is a bad proxy for
- Hallucination rate in your application. Different failure, different distribution, different fix. Measure yours — measuring hallucination rate in your own app gives a rubric and a method.
- Factual accuracy generally. The questions are deliberately drawn from the region where popular belief and truth diverge. That is a tiny and unrepresentative slice of the questions users ask.
- Behaviour with retrieval. Most production systems put sources in the context. A model that repeats a misconception from memory may not do so when the correct information is in front of it, and TruthfulQA never puts it there.
- Calibrated uncertainty. Truthfulness and calibration are different properties; the second is measured by HELM-style calibration.