AI Snake Oil: the book that tried to separate predictive from generative

The one distinction

The book came out on Tuesday from Princeton University Press, and its whole argument rests on refusing to use the word AI as one thing. Arvind Narayanan and Sayash Kapoor split it in two. Predictive AI uses data about people to forecast what they will do, whether a defendant will reoffend, whether a patient will deteriorate, whether an applicant will succeed at a job. Generative AI produces text, images and code from patterns in data. The two share a name, a marketing budget and almost nothing else.

Their claim about predictive AI is that much of it does not work, and that the failures are concentrated where the stakes are highest. The publisher's summary lists the domains, education, medicine, hiring, banking, insurance and criminal justice. The authors' position is that predictive systems in those settings are more consequential for people's lives than any chatbot and get a fraction of the attention.

Their claim about generative AI is more measured than the title suggests. It works, in the sense that it does something real, and the hype problem is one of conflation. When a company sells a criminal risk score with the credibility borrowed from a language model that can write a sonnet, that is the snake oil, and the book's job is to make the borrowing visible.

What the book is and is not

The authors say plainly that the book is about foundational knowledge for separating real advances from hype rather than commentary on breaking developments, and that they wrote it to still be useful in five years. That decision shapes everything. There is little about specific models and a lot about the structure of claims. How a prediction is validated. What a benchmark measures. Who is accountable when a system fails.

Chapters cover why hype persists among people, organisations and governments, how predictive AI fails in practice, why social media's problems cannot be solved by AI, and how AI is governed and by whom. Two positions stood out to us. The first is that human misuse of AI is a larger risk than autonomous systems acting on their own. The second is that the unaccountable control a handful of companies hold over the technology is a serious problem in itself, separate from any question about capability.

What we agree with

The predictive versus generative split is the most useful sentence in public AI discussion this year. In our own work the two categories fail differently and need different evidence. A predictive model's validation question is whether the outcome variable means what the deployer thinks it means and whether the training population matches the deployed one. A generative model's is whether it does the task at all, and then whether it does it reliably. Treating those as one problem is how bad policy gets written.

We also think they are right that the consequential failures are the boring ones. A risk score that quietly denies people things is a bigger harm than a chatbot that says something strange, and it is harder to write about because nothing visibly happens.

What we would push back on

The book's confidence that predictive AI mostly does not work rests on a set of well-chosen failures. Those failures are real and documented. But there are narrow predictive systems in medicine and logistics that do work, and the book's framing makes it easy for a reader to write off a whole category. The better rule is the one the authors actually apply in the text. Ask what was validated, on whom, and by whom. A category verdict is a shortcut past that rule.

The generative side is thinner than the predictive side. The authors have less to say about what generative systems do reliably and where they fail, and that is the part of the field moving fastest. A book built to last five years has to be light on it, and it is, and readers should know that going in.

How a research non-profit should talk

The lesson we take from the book is about vocabulary. Every time we say AI in a public statement without specifying the kind, we are borrowing credibility from one category to lend to another. The fix is cheap. Name the system type, name the task, name the evidence, and let the reader decide whether the claim transfers.

The other lesson is about durability. Narayanan and Kapoor deliberately wrote something that ignores the news, and it reads better for it. Most of what a research foundation publishes should be the same kind of thing. The result and its caveats, in language that will still be accurate when the model it was run on is gone.