Tests show that newer ChatGPT models hallucinate significantly more than older versions. The o3 model has a 33% hallucination rate, while the o4-mini hallucinates 48% of the time, compared to previous models (o1, o3-mini) which had 14-16% rates. OpenAI doesn't know why it's happening TechCrunch and says more research is needed.
Discussion
No replies yet.