The argument is stated early and it is a reasonable one. Every previous information technology, from writing to the printing press to radio, was a conduit. It moved information between humans, and humans did the deciding. This one, he says, is different in kind, because it can make decisions, generate content and act without a person in the loop, which makes it the first information technology that is a participant in the network rather than a channel within it.
That is a defensible thesis and a good organising idea for a book. The historical chapters are the strongest part and they are the majority of the book. The development of written records and what bureaucracy did to human society. How religious canons were assembled and the consequences of freezing a text.
The printing press and the fact that its early decades produced an enormous surge in witch hunting alongside the scientific revolution, which is a useful corrective to anyone who assumes that more information distribution automatically improves matters. The evolution of the press, of mass media, and of the institutions built to check them. This is popular history done competently and it is interesting whether or not you care about the AI argument. One idea in here is genuinely good and I have found myself using it since.
Information networks are not optimised for truth. They are optimised for order, meaning for coordinating large numbers of people, and truth is only sometimes useful for that. A story that is false but binding will beat a story that is true but divisive, over and over again, across every technology in the book. Apply that to a recommendation system trained on engagement and you have a sharper account of what went wrong with social media than most of the dedicated books on the subject manage.
Now the problem, and it is a serious one for a book that is marketed as being about artificial intelligence. The AI chapters are the weakest chapters. The description of how these systems work is vague at best and wrong at points. Terminology is used loosely, capabilities are described in ways that would not survive contact with anyone who has trained a model, and there is a persistent tendency to treat the systems as having intentions in a way that reads as sloppy rather than metaphorical.
For a book whose entire thesis rests on a claim about what distinguishes this technology from previous ones, that thinness is not a minor flaw. The method has the weakness it has had in his earlier books. A historical episode is described, an analogy is drawn to the present, and the analogy is then treated as having established something. History does not work like that.
The witch hunts and the recommendation algorithm may rhyme, and rhyming is not evidence, and there is a lot of weight resting on resemblance in this book. Length is a real issue. It is around five hundred pages and it contains perhaps two hundred and fifty pages of distinct material. Points are made, then restated with a fresh example, then restated again in summary, then recalled in a later chapter.
An editor with more authority would have produced a considerably better book. There is also a persistent grandness of tone that flattens distinctions. Sweeping declarations about what humanity has always done, or about the nature of information, delivered with total confidence and covering ground where specialists disagree sharply. It reads well and it does a lot of unearned work.
My two point nine is for a well written popular history with one genuinely useful idea, badly let down by the fact that the technology in its subtitle is the thing it handles worst. If you want the history of information networks, this delivers that. If you want to understand artificial intelligence and what it might do to society, the technical vagueness makes the conclusions impossible to trust, and you would be better served by a book whose author can accurately describe the thing they are worried about.