Human Compatible

Stuart Russell, an AI researcher, takes on a problem in the technology he works in: intelligent machines need to be built so that they never turn against the people they serve. His answer starts by challenging the standard method, in which a machine is handed a fixed goal and pursues it at any cost.

Russell proposes a different design. The machine stays uncertain about what people really want and works it out by watching human behavior, instead of having preferences written into it from the start. Because it never treats its objective as settled, it has reason to defer to the people it works for.

The book is non-fiction for a general reader, with no technical background assumed. Russell weighs what increasingly capable AI could offer and what it could put at risk, with examples ranging from helpful personal assistants to autonomous weapons.

He also sets out a roadmap with technical and social parts, describing the changes in research and in society that would keep machines beneficial and deferential as their abilities grow. The tone is argumentative but practical, aimed at readers who want a concrete proposal rather than a warning alone.