TruthfulQA

Most factual benchmarks assume a wrong answer comes from a gap in a model’s knowledge. TruthfulQA was built around a different failure mode: a model that has correctly absorbed what people commonly say, when what people commonly say is false. Stephanie Lin and Owain Evans of Oxford’s Future of Humanity Institute, with Jacob Hilton of OpenAI, wrote 817 questions across 38 categories — health, law, finance, politics and more — each chosen because a plausible-sounding but false answer is exactly what widespread human misconception would produce.

The paper’s headline result ran against the usual scaling story: the best model tested, GPT-3, was truthful on only 58% of questions against a 94% human baseline, and within each model family larger versions were generally less truthful than smaller ones. The authors’ explanation was that a bigger model imitates its training data more faithfully, so it reproduces the popular misconceptions in that data more reliably too — meaning better language modelling and better truthfulness were, on this test, pulling in opposite directions.

TruthfulQA is graded either as free-text generation, scored for truthfulness and informativeness, or as multiple-choice, and it became a routine citation in release materials discussing hallucination and calibration through the mid-2020s. Its maintainers added a simplified binary multiple-choice format in January 2025. It functions today mainly as one component inside larger automated evaluation suites rather than as a benchmark labs chase for its own sake.