The Problem of Apparently Irrational AI

Contemporary AI confronts us with a puzzle: how can such impressive language models behave in such apparently irrational ways? Social scientists have long struggled with a similar puzzle when studying other cultures: how should we understand strange behavior—and is it ever appropriate to regard it as irrational? In this paper, we argue that pragmatic, pluralist approaches developed in the social sciences to address these issues are useful for understanding language models. Moreover, we contend these approaches help us grasp when it is appropriate to regard behavior as irrational and better explained in non-rational terms.