Overview

• In a new paper and blog post, OpenAI calls hallucinations a persistent issue for large language models that cannot be completely eliminated.
• Researchers cite a chatbot giving multiple, conflicting false answers about author Adam Tauman Kalai’s Ph.D. dissertation title and birthday to illustrate confident errors.
• The paper links some errors to next-word pretraining that lacks true or false labels, making rare factual details difficult for models to predict reliably.
• OpenAI proposes updating evaluations to penalize confident wrong answers and grant partial credit for appropriate abstention, comparing the approach to tests that discourage blind guessing.
• The authors warn that widely used accuracy leaderboards must change or models will keep learning to guess, and they highlight tests showing lower error rates when systems refrain from answering uncertain questions.