Joy Buolamwini discovered as a graduate student that facial analysis software did not reliably detect her face, and that a white mask worked where her own face did not. That experience led to the Gender Shades research, which measured error rates across commercial facial analysis systems and found dramatically worse performance for darker skinned women. The findings changed products at several major companies and shaped policy discussions internationally. This book tells that story from the inside, which is what separates it from the many books about AI harm written by people who studied it from a distance.
The methodological account is the part I would press on practitioners. How you construct a balanced evaluation dataset, why aggregate accuracy figures conceal group level failure, and how to design an audit that a company cannot dismiss are all explained clearly and without jargon. Anyone who has ever reported a single accuracy number for a model should read this section, because it demonstrates concretely why that number can be actively misleading. That lesson generalises far beyond facial recognition.
The account of what happened after publication is unusually valuable. Research does not translate into change automatically, and the book describes the corporate responses, the pushback, the disputes over methodology and the eventual product changes and moratoriums. For anyone who wonders whether this kind of work matters, the answer is documented here, along with an honest picture of how slow and contested the process is. The memoir framing is a deliberate choice with real costs and benefits.
It makes the book readable for an audience that would never open a research paper, and the personal story gives the findings emotional weight that statistics alone do not carry. It also means a reader wanting only the technical substance will find themselves in extended passages about the author's life and career. I found the balance defensible and I understand readers who want more research per page. Scope is narrower than the title implies.
This is substantially a book about facial recognition and biometric systems, and the broader questions about language models, recommendation systems and automated decision making get much less attention. That focus is a strength in terms of depth and it does mean the book does not cover the full territory a reader might expect from the framing. Currency is a normal problem for a book in this field. Several of the corporate positions described have shifted again since publication, some products were withdrawn and some restrictions have loosened.
The specific corporate details date faster than the underlying analysis, and the methodology and the argument about measurement remain entirely current. Technical depth is moderate by design. A machine learning practitioner will want more detail on the evaluation protocol and the statistical treatment than the book provides, and the original papers are available for that. What the book supplies instead is the reasoning and the context, which is the part that is harder to get from a paper.
Four point zero. An authoritative account of algorithmic bias from the person who produced the evidence, written accessibly enough to reach a general audience without losing the substance. The measurement lessons alone justify the time for anyone building systems that make decisions about people, and the memoir structure is the price of that reach.