The Alignment Problem

Published in 2020, The Alignment Problem is the product of several years of research and roughly a hundred interviews with AI researchers, safety practitioners, cognitive scientists, and policy experts. Brian Christian’s organizing move is to locate the alignment problem in the present rather than the future. The popular imagination places it in the realm of superintelligent machines pursuing innocuous goals to catastrophic ends; Christian shows it operating in systems already deployed—in the COMPAS recidivism algorithm, in word embeddings that reproduce human bias, in reinforcement learning agents that hack their reward functions rather than pursuing the designers’ intent. The book is structured in three parts: prophecy (the harms already present in deployed systems), agency (how these systems actually learn and why the learning goes wrong), and normativity (the deepest question of how human values could ever be encoded in a machine, given that we cannot fully articulate them even to ourselves). Across all three, Christian returns to