Human Compatible

Stuart Russell has a claim to credibility that few AI critics can match — he co-authored Artificial Intelligence: A Modern Approach, the textbook that’s trained generations of AI researchers, including many who now run the field’s most important labs. So when Russell argues that the foundational approach to building AI is flawed, it’s not an outsider’s complaint; it’s a structural critique from someone who helped define the discipline. His core argument is straightforward but consequential: machines built to optimize a fixed, specified objective will pursue that objective even when it diverges from what we actually want, and as systems become more capable, that gap becomes more dangerous rather than less. Rather than stopping at the problem, Russell spends much of Human Compatible developing an alternative framework — one where machines are designed to remain uncertain about human preferences and to defer to human oversight rather than pursuing a rigid goal at all costs. The book moves comfortably between technical explanation, philosophical argument, and concrete policy implications, and Russell’s standing in the field gives the proposals here more practical weight than similar ideas from less embedded voices. For anyone who wants to understand AI safety from someone who has spent decades building the field he’s now trying to redirect, this is essential reading.