In this paper, I intend to analyse the relationship between AI and video games from a semiotic point of view. In the first part of the paper, I will propose a typology of artificial intelligence systems used in gaming software, based on their general functioning and on the parameters that govern their actions. I will start from the theory of enunciation formulated by Émile Benveniste, in order to define the relationship between virtual actions and the agency of both human and computational entities. Although Benveniste developed his theory to analyse verbal utterances, it was later adapted for application to other semiotic utterances, including still images, audiovisual forms, and video games. Extending the latter line of study, I will propose a typology of computational enunciation in video gaming and identify four fundamental trends in the functioning of modern games: script-based AI, reactive AI, modelling AI, and co-enunciating AI. Finally, I will focus on the fourth of these in order to analyse how AIs are used within video games with the aim of producing responsibility and empathy effects towards secondary characters.