A different way of thinking about artificial intelligence, one that will maximize human values and lead to provably beneficial systems, from a leading AI researcher In the popular imagination, superhuman artificial intelligence is a terrifying inevitability. It's not just that robots could take over many jobs; machines could deviate from their programmed purpose or be imbued with objectives that are detrimental to our own. In this groundbreaking book, distinguished AI researcher Stuart Russell argues that this scenario is not destined to happen-as long as we make sure that the purposes we put into machines are the ones we really desire. Russell begins by reviewing the past, present, and future of our attempts to understand and create intelligence. He describes the AI breakthroughs still to happen, from intelligent personal assistants to machine systems capable of cumulative learning, and spells out the ways humans will find to misuse AI, from lethal autonomous weapons to viral sabotage. We could face grave threats from AI systems that are more powerful than humans. How do we deal with this challenge' In Human Compatible, Russell suggests that we may need to partially reshape the foundations of AI to provide for machines that pursue our objectives, not theirs. Such a machine would be a humble, altruistic one that is unaware of its true purpose. His solution focuses on creating uncertainty in the robot's objectives and teaching them to fill that gap with knowledge of human values learned through reading human texts and observing human behavior. Superhuman artificial intelligence may arrive considerably earlier than expected; when and if it does, it would be the biggest event in human history. Getting human objectives aligned with those of machines will be the key to harnessing the full power of AI while also preventing the Armageddon of robot takeover.