Generative Artificial Intelligence (GenAI) and a Taxonomy of Undesirable Results

ABSTRACT This paper makes the case that the term “hallucinations” is inadequate for describing undesirable results from Generative Artificial Intelligence (GenAI) output. We explain hallucinations, how the term is used, and why it is inadequate. We propose a taxonomy for categorizing these undesirable results, facilitating clearer communication about them. This taxonomy aids in understanding, troubleshooting, and developing resilience against undesirable results, which is important for corporate governance as GenAI integrates both specifically into enterprise resource planning systems, and also software, generally. We give several examples illustrating how our taxonomy provides the necessary language to describe and address undesirable results. We discuss the implications of using the taxonomy widely, particularly in the context of text-to-text Large Language model output, and consider its relevance across different GenAI modalities, as common GenAI multimodal solutions expand.