Uncertainty

Abstract

Abstract The arguments in preceding chapters leave significant uncertainty about whether AIs have welfare. This uncertainty stems from three sources. First, there is uncertainty about what is required for welfare subjecthood: Is it agency, consciousness, or sentience? Second, there is uncertainty about what it takes to have beliefs and desires, consciousness, pleasure, and displeasure. Third, even once it is known what it takes to have these properties in general, there is uncertainty about whether particular AIs instantiate them. For example, there may be persistent uncertainty about whether LLMs implement a global workspace architecture somewhere inside their neural nets. In the face of this uncertainty, how should one treat AIs? This chapter argues that humanity should already give moral consideration to AIs in decision-making because moral decision-making must be sensitive to uncertainty.