[mfoc]

The term 'hallucinations' is by now a common expression when speaking about the shortcomings of LLMs and Generative AI. However, I propose that the expression is misdirected at LLMs and takes away rather than adds to our understanding of the underlying issue.

It is humans who hallucinate when they expect an answer from an LLM that is 'truthful' or 'right' or 'correct'. The LLMs (as of June 2024) have no concept or understanding of truth, falsehood, right, wrong, correct or incorrect. Rather, when people claim that LLMs hallucinate, the LLMs are in fact generating responses and performing exactly as they are programmed to do.

However, humans should be forgiven for their hallucinations, because the interaction with the LLMs and the generated results are often designed by their makers to precisely give this unfortunate impression. Indeed, it is the unreasonable effectiveness of LLMs in the natural language domain and their successful application to many diverse tasks, that 'hallucinations' are now seen as the most serious defect of LLMs.