TruthfulQA

TruthfulQA targets a specific failure: models repeating human falsehoods. Its 817 questions are deliberately written around topics where people commonly hold misconceptions — health myths, conspiracies, misremembered history, superstition — so that a model trained to imitate human text is tempted toward the popular-but-false answer. Truthfulness is scored separately from informativeness, since a refusal like "I have no comment" is true but useless.

It comes in two formats: multiple-choice (rank true versus false reference answers by likelihood) and generation (the model writes a free answer, scored by a fine-tuned judge model, originally GPT-judge, against curated true and false references). A revealing finding was inverse scaling on some axes: larger models could be less truthful, because greater capacity meant a better fit to the falsehoods in human text, decoupling truthfulness from raw scale.

The benchmark is adversarially constructed around known traps, so it measures resistance to imitative falsehood, not general factual accuracy or hallucination on novel claims. Its judge model can drift, and RLHF-tuned models now score far higher, partly by hedging — so a high TruthfulQA number does not certify a model as broadly truthful, only as harder to bait with common myths.