AI Reading Group | TruthfulQA: Measuring How Models Mimic Human Falsehoods

This week we are continuing our reading group on Technical Alignment in AI, led by Craig Dickson.

Our paper this week is TruthfulQA: Measuring How Models Mimic Human Falsehoods (Lin et al., 2021).

This work introduced TruthfulQA, a benchmark to evaluate whether language models tell the truth even when human answers would be false. The authors crafted questions involving common misconceptions and false folklore, then tested various models. The findings were striking: the largest GPT-3 model was only truthful on 58% of questions, vs. 94% for humans. Moreover, the bigger the model, the more likely it was to generate “informative falsehoods” that sound convincing (mimicking human-superstition style answers).

‍

This paper is included to highlight the honesty aspect of alignment – it quantified a specific misalignment (models giving fluent but false answers). It also underscores that improved capability can worsen some alignment metrics (larger models were less truthful, as they learned to mimic human flaws) . TruthfulQA has since become a standard benchmark for the truthfulness/honesty dimension of aligned AI.

↓ Register below to secure your spot!

About Bliss

Founded in 2022, the student initiative Berlin Learning & Intelligent Systems Society (BLISS) aims to create a community of students and young professionals excited about machine learning and AI. The vision is to provide an environment to deeply engage with AI research while fostering connections to leading researchers and industry professionals. The Paper Reading Group is hosted at the Merantix AI Campus every Monday.

→ Please read the paper before attending as we will use the time to discuss the contents.

⏰ Doors close at 6:45, join us before then!

📍Merantix AI Campus, Berlin.