Machines are no longer only tools. They are becoming delegates, trusted to exercise judgement no one can fully inspect.

Drawing on philosophy of mind, neuroscience, and the technical record, this book separates what has actually been observed in AI systems from what has been extrapolated. It engages Nick Bostrom’s orthogonality and instrumental convergence theses, tests Eliezer Yudkowsky’s founding arguments against two decades of evidence, and asks why values that are formed may hold where values that are installed give way.

Neither a counsel of despair nor a reassurance, it is an account of what the alignment problem is, why it is hard, and what it would take for human beings to remain the locus of judgement in a world of increasingly capable machines.