Primary category
The Alignment Problem
Machine Learning and Human Values
Machine-learning systems learn from examples, rewards, and human behavior—but none of these provides a simple or neutral expression of what people truly want. Brian Christian investigates the resulting alignment problem: the difficulty of making computational systems reliably serve human purposes and values. Moving between the history of artificial intelligence, psychology, moral philosophy, and contemporary research, he examines biased representations, incompatible definitions of fairness, opaque models, poorly specified rewards, imitation learning, preference inference, and uncertainty about human objectives. The book treats alignment not merely as a future problem involving hypothetical superintelligence, but as a present challenge wherever automated systems classify people or influence consequential decisions. Christian’s central contribution is to connect technical design choices with the unresolved ambiguities of human judgment. The result is an accessible account of why buildin…