Hannah Fry is a mathematician at UCL who has become one of the better science communicators working, and Hello World is her book about what happens when algorithms make decisions that matter. It is organised by domain: power, data, justice, medicine, cars, crime and art, and each chapter takes real deployed systems and asks what they actually do to the people they are applied to. The structural choice that makes the book work is the insistence on real cases. The justice chapter deals with recidivism prediction tools used in American sentencing and parole decisions, and rather than treating them as an abstract fairness puzzle it walks through what it means for a specific person to receive a longer sentence because a model assigned them a risk score, what the model was actually measuring, and what the competing definitions of fairness would each require.
That is much harder to wave away than a trolley problem, and much more instructive. The medicine chapter is the one I find most valuable, because it makes a point that technical audiences consistently miss. An algorithm that outperforms the average doctor is not automatically an improvement to healthcare, because the question is what happens at the interface: how clinicians interpret a score, whether they can override it, whether they stop developing their own judgement, whether the system's failures are correlated in ways an individual's are not. The performance number is the beginning of the analysis rather than the end of it.
Fry's central framing is where humans should stay in the loop, and she is consistently more interested in the partnership than in the horse race. Her answer is usually that algorithms are good at consistency and scale and bad at context and exception, humans are the reverse, and the design question is how to combine them rather than which to pick. That is a more productive frame than most algorithmic ethics writing manages, and it is directly applicable if you build these systems. She explains statistics well, which is the technical spine of the book.
Base rates, false positive and false negative trade-offs, why a test that is 99 percent accurate for a rare condition still produces mostly false alarms. These are the concepts people get wrong when reasoning about model deployment and she teaches them cleanly without either condescending or retreating into notation. The limitation is timing. Published in 2018, it is a book about classical algorithms and predictive models, not generative AI.
The examples are risk scores, recommendation engines, facial recognition and autonomous vehicles. That means it misses the entire current wave, and some of the specific systems discussed have since been modified or withdrawn. What it does not miss is the structure of the problem, because the questions of accountability, contestability, appropriate deference and who bears the cost of an error are exactly the questions now being asked about generative systems, and they were better articulated here than in most of what has been written since. Depth is traded for breadth throughout.
Seven domains in three hundred pages means none is exhausted, and a reader wanting the technical detail of any one will need to go elsewhere. Four point zero. Excellent, humane, genuinely useful on the questions that matter, dated in its examples and not in its thinking.