More problems of large language models in comparison with human knowledge

Abstract This paper builds on an earlier paper in this journal (Collins’s, 2024, ‘Why artificial intelligence needs sociology of knowledge …’). In that paper, an attempt was made to explain why LLMs had no moral compass. Here, we first explain more carefully where humans get their knowledge from and fill out the problems of LLMs by comparing them with humans. We then provide a new classification of different ways in which LLMs fail. There is one main cause, which is to do with sources of information, and there are three sub-categories. The first is to do with failures of reflexivity and has three sub-divisions; the second relates to the way LLMs interact with humans and has four sub-divisions; the third is to do with simple technical mistakes and has six sub-divisions. Each of the 13 kinds of problem is illustrated with an example of failure. We hope this new classification is exhaustive, but invite additions if it is not.