Abstract

Large Language Models are one of the most prominent examples of current uses of AI, and one of the most urgently pursued normative tasks is to make them and their interactive user-interfaces—open-domain chatbots—safe. However, in this paper, we elaborate first on why such a limited view on the permissibility and desirability of their utterances falls conceptually flat, is philosophically insufficient, and leads to severe technological limits. We then propose a positive normative concept, appropriateness, that can provide the required orientation for determining what chatbots may or may not say, and distinguish between technical-discursive, social, and moral appropriateness. Lastly, we explore the requirements for social and conversational contexts to fulfill appropriateness-requirements, and find that positionality is a key element for their position in discourse.