This is not one author's book: these are lectures from an advanced course, grouped into five parts — human-centred machine learning, explainability, ethics and law, argumentation, and social simulation. Like every collected volume, the quality rises and falls from chapter to chapter.
What I took from it
The ethics part is the one that actually helps, and not for its declarations. One chapter is devoted to why abstract ethical principles do not apply themselves, and another to how you carry out a socio-technical assessment of a specific system. The gap between those two things is enormous: a list of principles gets signed and filed; an assessment asks who answers for a wrong output, where the complaint goes, and what happens while it is being sorted out.
From there I took the question I now ask whenever someone wants a model deciding about people or about money: which decision stays with the system, which stays with a named human being, and how that person finds out they need to step in. If those three answers do not exist, there is no system, there is an experiment running on real users.
The explainability chapters left me something similar: explainable is not a property of a model, it is a relationship with someone. Explaining to the person who writes the code, to the person who audits it and to the person on the receiving end of the decision are three different jobs, and conflating them is the most common way to deliver an explanation nobody can use.
What I left out
Almost all of the first part, a fairly standard deep learning course, plus the interactive robotics and social simulation sections, which are far from what I do. There is even a chapter on writing science fiction as a way of disseminating research: charming, and of no use whatsoever in my trade.