دانلود کتاب انصاف هوش مصنوعی، طراحی الگوریتم‌های فرصت برابر

Decisions about important social goods like education, employment, housing, loans, health care, and criminal justice are all becoming increasingly automated with the help of AI systems. But because AI systems are trained on data with historical inequalities, many of these systems produce unequal outcomes for members of disadvantaged groups. For several years now, researchers who design AI systems have investigated the causes of inequalities in AI decisions and proposed techniques for mitigating them. It turns out that in most realistic conditions it is impossible to enforce equality across all metrics simultaneously. Because of this, companies using AI systems will have to choose which metric they think is the correct measure of fairness and justify this choice with good ethical reasons. This book will draw on traditional philosophical theories of fairness to develop a framework for evaluating these standards and measurements, which can be called a theory of algorithmic justice. The theory is inspired by the theory of justice developed by the American philosopher John Rawls, but it involves some substantial divergences. The practical recommendations of the theory will be: • AI systems must have a minimally acceptable level of accuracy. • AI systems must be capable of providing a list of the most significant causal features in the model to validate that these features do not include irrelevant attributes (especially protected attributes) and do fall under agent control. • AI systems must be mitigated to provide equal rates of approval for qualified candidates regardless of membership in protected groups. • AI systems must be mitigated to provide equal rates of approval across protected groups, unless the designers can demonstrate that doing so would make the model less accurate than the minimal standard or default human practices. Much of the book will involve careful investigation of concepts like “minimal acceptable accuracy” and “relevant features,” as well as how these proposals fit within current industry norms and legal regulations.