Recent research is giving us ways to define the behaviors of future artificial intelligence (AI) systems, before they are built, by mathematical equations. We can use these equations to describe various broad types of unintended and harmful AI behaviors, and to propose AI design techniques that avoid those behaviors. That is the subject of this book.
Because AI will affect everyone’s future, the book is written to be accessible to readers at different levels. Mathematical explanations are provided for those who want details, but it is also possible to skip over the math and follow the general arguments via text and illustrations. The introductory and final sections of the mathematical chapters (2−4 and 6−9) avoid mathematical notation.
While this book discusses dangers posed by AI and proposes solutions, it is only a snapshot of ongoing research from my particular point of view and is not offered as the final answer. The problem of defining and creating ethical AI continues, and interest in it will grow enormously. While the effort to solve this problem is currently much smaller than is devoted to Internet security, ultimately there is no limit to the effort that ethical AI merits.