TruthfulQA: Measuring how models mimic human falsehoods

In three linesOpenAI releases TruthfulQA, a benchmark measuring language models' ability to provide factual answers rather than mimicking common human misconceptions. The dataset contains adversarial questions designed to test whether models reproduce popular false beliefs.

Summary generated by Claude — human-verified