Mapa das Instalações

Abstract: With the rising capability of AI systems, the field of AI safety argues that we are likely to face issues with aligning future AIs with human values. These fears are motivated by the behaviour of current AIs, which often misinterpret their given objective. However, theorists of this field tend to model AIs according to the axioms of instrumental rationality, which portrays the agent's desires as something given and fixed, often arguing that behaviour is motivated by the agent's utility function or simply a final goal. On these grounds, these theorists go on to argue that what is necessary is to specify a harmless final goal to align AIs. In this paper, I argue against this approach because it creates a dualism between inner values and the outer world, which prevents the modelling of embedded agency. Embedded agents are likely to deal with radical uncertainty in their given environments and may end up changing their goals in unexpected ways. Because future AIs are likely to be such agents, what is necessary is to formulate a different conception of rational agency to solve the alignment problem. Drawing inspiration from pragmatism and 4E cognition, I will argue that we need to model agency based on axiological reason, which is concerned with the appropriateness of ends given the circumstances. To deal with AI, we cannot merely equip it with 'correct' goals but rather with the right process of selecting its ends and means.