The Alignment Problem: Machine Learning and Human Values
Decide its fate
Ask the word-embedding model to finish an analogy and it answers "doctor - man + woman = nurse," a line one reader copied into his notes. Christian skips the paperclip-apocalypse routine and argues from harms already on the record: the parole algorithm that overpredicts reoffending for Black defendants, the film stock calibrated to a white model's skin, since "Darker skin colors were completely ignored by manufacturers of cameras." It lands at both ends of the room. A reader who calls herself more farm than lab kept one researcher's confession, "I always felt that what I really wanted to learn was to think well," while a specialist fluent in the alignment literature credits it with "doing justice to the topics I knew well," a test he says most popularizations flunk.
Four hundred ninety-six pages, and the question on the cover, whether machines can learn human values, never gets answered. What fills the space instead: anecdote after anecdote of systems misbehaving, each expert wheeled in with a full academic biography first, "the countless stories with tiny variations to prove one single point." Meanwhile the jargon arrives undefined. One reader: "I felt like I missed the prerequisite course that was required for this book." Another goes further and warns, "Only read this if you have taken a graduate-level course on machine learning," a strange requirement for a book sold as the layperson's on-ramp to the field. One reviewer offered the publisher a fix: "To make this not clickbait, delete 80% of the pages or fix the title."
