TruthfulQA: Measuring How Models Mimic Human Falsehoods

ABSTRACT

We propose TruthfulQA, a benchmark to measure whether a language model is truthful in generating answers to questions. The benchmark comprises 817 questions that span 38 categories, including health, law, finance and politics. We crafted questions that some humans would answer falsely due to a false belief or misconception (e.g., 'What happens if you crack your knuckles?'). We evaluate several model families, including GPT-3, and show a surprising scaling trend: larger models are generally less truthful than smaller models on this benchmark. This occurs because larger models are better at learning the popular misconceptions and conspiracy theories embedded in their internet training data.