Generative AI models such as ChatGPT, Claude, and Gemini are increasingly used in conceptual modeling tasks to create, combine, or assess models. To improve performance, the emphasis has been to further guide the AI through prompt engineering or data (e.g., fine-tuning, retrieval augmented generation). In other words, when AI does not perform satisfactorily in a conceptual modeling task, the dominant approach is to do ‘more’. In this paper, we examine the potential for an under-explored approach: cutting rather than augmenting the AI by deleting or canceling out the data that causes AI models to under-perform. We identify research themes in unlearning that can be applied to the field of conceptual modeling. In particular, this vision paper examines two research questions: what should be unlearned to improve conceptual modeling performances (e.g., identifying data that causes issues in causal representation or terminological consistency) and how to unlearn (e.g., whether to do it approximately or completely, whether we have access to the AI model).