Big insight from OpenAI’s recent research: Hallucinations in LLMs aren’t just bugs—they stem from how we train & evaluate models, rewarding confident answers over admitting uncertainty.

Their paper shows why models guess boldly instead of saying ‘I don’t know,’ but the good news? Newer models are already dropping rates significantly (GPT-5 family seeing major gains in reasoning truthfulness). This transparency is huge—it’s guiding better evals & techniques to build more reliable AI. 

openai.com/index/why-la…