A note on ethics and prediction

Drawing on the work of Guglielmo Tamburrini, we connect his principle of ‘expected perception’ to the debate on AI ethics. This principle, which models how an AI agent learns by minimizing the error between its predicted and ac- tual sensory inputs, offers a powerful lens for understanding AI autonomy. We argue that this mechanism, rooted in predictive processing, provides a plausible path to intrinsic and fair value alignment. By learning to predict and minimize social prediction errors, the agent could be motivated to act ethically, seeing moral behaviours as statistical regularities to be upheld. This perspective offers a fresh view for building inherently beneficial AI.