This Book provides a layman's perspective on algorithmic decision making and social values like Privacy, Fairness, Transparency and Interpretability, Morality, and Safety. It is a highly recommended read for anyone who wishes to understand the pitfalls of Artificial Intelligence and Algorithmic decision making. The best part is that the authors have explained complex mathematical and probability concepts in a layman's language so that even a social science reader can understand it well.
Each chapter is dedicated to a specific theme. For instance, the first chapter on privacy discusses the pitfalls of privacy conversation. It explains how solutions like anonymization are incomplete. It gives examples of Arvind Narayan's Netflix research, legal pitfalls of movie database records, etc. It is interesting to know that despite anonymization, any person can be zeroed upon with 6 geolocations in the entire day. It argues about the issue of privacy vs predictive accuracy and explains the solution of using differential privacy as a regulator knob.
Similarly, the chapter on fairness is filled with numerous examples. The book is engaging in its analysis and explanations. It has tried to bring out both sides of the arguments. W.r.t. Fairness, the book has beautifully explained the challenge of describing what is fairness and synthesize the entire discussion in form of pareto curve and pareto frontier. The definition of fairness as equality of false positives or euqality of false negatives is best explained using this. In this, author also highlights that science can provide the trade offs between different definitions of fairness but ultimate decision would hinge on human wisdom.
One chapter was dedicated to algorithmic game theoretic where authors explain issues in prisoner dilemma. It explains how technological solutions like Waze and Google Maps are promoting self interest over social welfare or competitive equillibrium. It also highlights other issues like eco chammberedness in machine learning and recommendation engines. Further suggestions like Cooperation through correlation are given as example.
Book also provided insight into email scams and how adaptability and scale is being used to fool people about market prediction. The concept of p-hack and garden of forking paths highlights the issue of false scientific research in the community.
Overall the book is an interesting read with umpteen examples, simple lucid language and open ended perspectives. A must read for folks who wish to know about various pitfalls of algorithmic processing and AI-ML.
