What ought AI do?

This article draws upon Georg Simmel’s engagement with the concept of ought as a resource for exploring the question of what it is that AI ought to do. In so doing, it draws upon classical sociology to critically explore and reframe a defining question in AI development with regard to the identification and setting of ideal or expected behaviours of AI models. In particular, Simmel’s relational and contingent conceptualisation of ought in his final book The View of Life informs an exploration of how the ought is prescribed and authorised in the development of AI. Applying this perspective, the article looks at how the ought is implicitly demarcated in the development of AI models, and, further to this, looks at an example of how the AI sector seeks to find and then establish collective ideals of expected AI behaviours. Inspired by Simmel’s late work, the article focuses upon how AI is modelled around a version of the ought. It then narrows this further by looking at how notions of ‘collective alignment’ operate to establish the ought, in turn changing the conditions of AI development and shaping the ongoing design of AI models themselves. The article is both a sociology of AI and also a sociology of the conceptual and methodological practices behind AI development. It draws upon Simmel’s conceptual ideas to show the problems with this external version of the ought as adapted in AI modelling, and why it is that sociology should ask the question: what ought AI do?