Bender and Hanna have been making this case for years, Bender from computational linguistics and Hanna from sociology and a stint inside a corporate ethics team, and the book collects the argument properly. The claim is that a great deal of what the industry sells as intelligence is marketing built on statistical pattern matching, that the language used to describe these systems does persuasive work the technology does not support, and that the costs in labour, data appropriation and environmental impact are systematically excluded from the accounting. The strongest chapters are on language. The authors are unusually good at showing how describing a system as understanding, reasoning, hallucinating or learning imports assumptions that the underlying mechanism does not justify, and how those word choices shape what buyers and regulators believe they are getting.
Once you have read this you notice it constantly, in vendor documentation, in press coverage and in your own writing. That alone justifies the time. The treatment of labour is the part the industry most needs to hear. The human work behind data labelling, content moderation and reinforcement from human feedback is real, poorly paid, often traumatic and almost entirely absent from how these products are described.
The book documents it seriously. Anyone who has read only vendor material about training data will find this uncomfortable, which is the point. Data provenance and environmental cost get similar treatment, with more rigour than the usual commentary. My criticism is the register.
The book is written with a sharpness that will read as bracing to sympathetic readers and as dismissive to anyone who works with these tools daily and finds them genuinely useful. The arguments are strong enough that they do not need the tone, and the tone gives the people who most need to engage an excuse to stop reading. I say this as someone who agrees with a good deal of the substance. The related weakness is that practical capability gets less credit than it has earned.
These systems do useful work. They summarise, translate, draft, extract structure from mess and shorten a great many tasks that used to take hours. Acknowledging that more fully would strengthen the critique rather than weaken it, because the interesting argument is about the gap between what they usefully do and what they are sold as doing, not about whether they do anything. What it does not offer is a way forward for practitioners.
If you build with these tools and finish the book persuaded, the guidance available is largely to be sceptical and to resist. That is a legitimate position for authors to take and it leaves the reader who has to ship something on Monday without much. A chapter on what responsible use looks like, given that adoption is happening regardless, would have widened the audience considerably. Read it anyway, and read it if you already agree, because sitting with a well made argument against your own work is the most useful hour you can spend.
Four point zero. Rigorous, necessary and rooted in genuine expertise, marked down slightly for a tone that will lose the readers who most need it and for critiquing without offering the practitioner anywhere to stand.