False Responses From Artificial Intelligence Models Are Not Hallucinations.

Abstract

As recently highlighted in the New England Journal of Medicine,1,2 artificial intelligence (AI) has the potential to revolutionize the field of medicine. While AI undoubtedly represents a set of extremely powerful technologies, it is not infallible. Accordingly, in their illustrative paper on potential medical applications of the recently launched large language model GPT-4, Lee et al. point out that chatbot applications for this AI-driven large language model occasionally produce false responses and that “A false response by GPT-4 is sometimes referred to as a ‘hallucination,’.”1 Indeed, it has become standard in AI to refer to a response that is not justified by the training data as a hallucination.3 We find this terminology to be problematic for the following 2 reasons: It is not constructive to merely criticize a terminology without providing an alternative. Therefore, given the topic and the timing, we sought advice from AI. Specifically, we first turned to GPT-3.5—the proverbial predecessor of GPT-4: