New OpenAI research cracks the case on why AI models hallucinate (and what to do about it)
Made with Midjourney
Large language models have gotten better at avoiding mistakes — but they still make things up. While it isn’t a big deal in some low-stakes scenarios, you probably don’t want an AI model to hallucinate in sensitive use cases like law and medicine. A new research paper from OpenAI digs into why even advanced models like GPT-5 still generate “plausible but false statements,” and found some interesting answers.
Surprisingly, advanced models are hallucinating more. On PersonQA, which questions the model on public figures and facts, OpenAI’s o3 model hallucinated 33% of the time, double its predecessor o1. The smaller o4-mini fared even worse at 48%. That’s despite both models showing stronger math skills. It’s a paradox: advanced models can be smarter and more wrong at the same time.
The key finding: Current training and testing methods penalize models for admitting uncertainty. These incentives create pressure for models to respond, even when they don’t know the answer.
OpenAI says benchmarks must change. Instead of rewarding only right answers, evaluations should penalize confident mistakes and give credit for saying “I don’t know,” like exams that dock marks for wrong guesses but not for leaving a question blank. Hallucinations won’t vanish, but with better incentives, models could learn to fail more honestly.