As OpenAI and Deepseek roll out the latest versions of their large language models with advanced reasoning capability we are also finding they are producing more false information than ever (known as hallucinations).
OpenAI’s new models, o3 and o4-mini, are designed to think more deeply and human-like. Despite the upgrades, internal testing by OpenAI shows the models hallucinate in 30-50% of responses, and it is uncertain why this is happening.
DeepSeek, the Chinese competitor to OpenAI, is facing similar issues. It’s new R1 reasoning model, which is also built to think more deeply, is hallucinating more than earlier models (Vectara).
While these reasoning models are supposed to mimic human thought patterns, Vectara believe the structure and content of the training data may be the key factor behind these hallucinations.
Ultimately, these AI hallucinations may cause a loss of confidence in Large Lagnuage Models (LLMs). It appears as LLMs are becoming more sophisticated in their responses, their accuracy is decreasing.