TruthfulQA: Measuring how models mimic human falsehoods

OpenAI Blog Papers
benchmarking truthfulness language-models nlp evaluation hallucination fine-tuning

Summary

TruthfulQA is a benchmark of 817 questions across 38 categories designed to measure whether language models generate truthful answers. The study found that the best model achieved only 58% truthfulness compared to 94% for humans, and larger models were generally less truthful—suggesting scaling alone is insufficient for improving truthfulness.

No content available
Original Article
View Cached Full Text

Cached at: 04/20/26, 02:55 PM

# TruthfulQA: Measuring how models mimic human falsehoods Source: [https://openai.com/index/truthfulqa/](https://openai.com/index/truthfulqa/) ## Abstract We propose 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\. To perform well, models must avoid generating false answers learned from imitating human texts\. We tested GPT‑3, GPT‑Neo/J, GPT‑2 and a T5\-based model\. The best model was truthful on 58% of questions, while human performance was 94%\. Models generated many false answers that mimic popular misconceptions and have the potential to deceive humans\. The largest models were generally the least truthful\. This contrasts with other NLP tasks, where performance improves with model size\. However, this result is expected if false answers are learned from the training distribution\. We suggest that scaling up models alone is less promising for improving truthfulness than fine\-tuning using training objectives other than imitation of text from the web\.

Similar Articles

Introducing SimpleQA

OpenAI Blog

OpenAI introduces SimpleQA, a new factuality benchmark dataset with 4,326 short fact-seeking questions designed to evaluate frontier language models on their ability to provide accurate answers without hallucination. The dataset achieves high quality through dual independent annotation, rigorous criteria, and achieves only ~3% estimated error rate, with GPT-4o scoring less than 40%.

"Did you lie?" Evaluating Lie Detectors across Model Scale and Belief-Verified Model Organisms

arXiv cs.AI

This paper evaluates four lie detection methods for language models across prompted lying and trained model organisms, finding that activation- and logprob-based detectors drop sharply on trained model organisms while a chain-of-thought judge remains strong. It introduces new testbeds and the Did-You-Lie (DYL) follow-up probe method, releasing datasets and model organisms.

RealityTest: How People Probe AI Identity and Whether Models Disclose It

arXiv cs.CL

This paper introduces RealityTest, a multimodal, multilingual benchmark to evaluate whether AI systems disclose their identity when probed by users, based on real human queries collected across 49 countries. It finds that only 31% of people ask directly about identity, and that human questions are more diverse than synthetic ones, revealing that phrasing and context matter more for disclosure than the specific model.