People ask me for a reading list several times a month, usually because they have been handed an AI strategy and given a fortnight to have an opinion about it. This is the list I send.
It is organised by what you need, rather than by how important the book is. Each entry says plainly what the book is for and who should skip it. I have written one of them, and I have put it where it belongs rather than at the top.
Two books that will give you the shape of the argument without requiring you to care about machine learning.
Arvind Narayanan and Sayash Kapoor · Princeton University Press, 2024
The most useful distinction in the field: predictive AI, which mostly does not work, and generative AI, which works unevenly. Two Princeton computer scientists, so the technical claims hold up.
Read it if: you need to tell a colleague why a procurement pitch is overstated, and you want a citation rather than a feeling.
Emily M. Bender and Alex Hanna · Harper, 2025
A linguist and a sociologist on how AI is sold, and on who absorbs the cost when it fails. Sharper in tone than Narayanan and Kapoor, and less interested in being fair to vendors.
Read it if: you want the critical case made at full strength. Skip it if you are looking for something to hand a sceptical senior manager.
Books written by people who have actually sat in a classroom, which is a smaller category than you would hope.
Justin Reich · Harvard University Press, 2020
Why education technology keeps promising transformation and delivering modest, uneven gains. Predates the current wave and is more useful for it, because the pattern it describes is the one repeating now.
Read it if: your institution is about to spend significant money on a platform that will fix learning.
Audrey Watters · MIT Press, 2021
The long history of automated teaching, from Skinner onwards. Establishes that almost nothing about the present argument is new.
Read it if: you want to know why the promises sound familiar.
Neil Selwyn · Polity, 2019
Takes the question seriously rather than dismissing it, and is better for that. Clear on what automation can and cannot carry in a classroom.
Read it if: you are being asked to justify why a human should do a task a machine could approximate.
Sam Illingworth · Self-published, 2026
Mine. Four tests for deciding whether a given task should go near a model, and what it costs you when the answer is no and you do it anyway. Ninety-nine pages, and short on purpose.
Read it if: you want something practical to use on Monday. Skip it if you want the research literature, because it is a short book and it cites rather than surveys.
Where the power, the money and the labour actually sit. These are the books that change how you read a vendor's sustainability page.
Kate Crawford · Yale University Press, 2021
Follows AI back to lithium mines, data labour and energy. The book that makes it impossible to treat a model as though it were weightless.
Read it if: you are writing anything about AI and sustainability.
Karen Hao · Penguin Press, 2025
Reported from inside the companies. The best account of how the decisions actually get made and by whom.
Read it if: you want to understand the institutions rather than the technology.
Cathy O'Neil · Crown, 2016
The original argument that opaque scoring systems entrench the inequality they claim to measure. Pre-dates generative AI entirely and remains the clearest statement of the problem.
Read it if: you are designing or buying anything that ranks students.
Ruha Benjamin · Polity, 2019
How discrimination gets built into systems that nobody intended to be discriminatory. Essential before you write an equality impact assessment for an AI tool.
Read it if: your institution is about to describe a system as neutral.
Read these alongside the enthusiastic books rather than instead of them. Arguing with a position you have not read is how people end up losing the argument in committee.
Ethan Mollick · Portfolio, 2024
The most influential optimistic account, and genuinely useful on practice. Light on the costs and on the cases where the honest answer is to not use the tool.
Read it if: you need to know what the people across the table have read.
Shannon Vallor · Oxford University Press, 2024
A philosopher's answer to the optimists: these systems reflect us rather than surpass us, and the mistake is treating the reflection as a destination.
Read it if: you found Mollick persuasive and want to know what is missing from it.
Brian Christian · W. W. Norton, 2020
The clearest explanation of why making these systems do what we mean is hard. Technical without being inaccessible.
Read it if: you keep being told alignment is solved, or that it is hopeless.
These are the books that come up when people ask me, and the ones I find myself recommending most often. Several of them contradict each other, which is the point. A reading list where every book agrees is a reading list assembled to confirm something.
There are no affiliate links on this page and I earn nothing if you buy any of these, including my own, if you borrow it from a library instead.
If you think something is missing, tell me. I update this page when I am persuaded.
I write Slow AI, a newsletter about when AI helps, when it does not, and how to tell the difference. It is read by more than 23,000 people in 179 countries, and one post a week is free.
Read Slow AI