The Means of Prediction and the Ownership of AI Objectives
Maximilian Kasy's The Means of Prediction: How AI Really Works (and Who Benefits) is a useful antidote to mystical AI talk because it makes the control question material. The book asks who controls AI objectives and the resources that make prediction systems possible.
For this review, the means of prediction are data, compute, expertise, and energy. Whoever governs those inputs can shape which errors are tolerated, which objectives become infrastructure, and whose uncertainty gets turned into someone else's product.
The Book
The Means of Prediction: How AI Really Works (and Who Benefits) was published by the University of Chicago Press in November 2025. The publisher lists Maximilian Kasy as the author, gives the print ISBN as 9780226839530, and lists the book at 224 pages. Amazon's listing confirms the retail path, ISBN-10, and ASIN as 0226839532.
The official table of contents is unusually direct. After an introduction, it moves through supervised learning, overfitting and underfitting, deep learning, and the exploration/exploitation trade-off. It then turns to machine power, value alignment, privacy, automation, fairness, explainability, and democratic control of the means of prediction. That sequence matters: Kasy does not separate technical explanation from institutional power.
The Core Argument
The book's strongest move is to refuse the simple humans-versus-machines frame. In this reading, the central conflict is not whether a machine will replace a person in the abstract. It is which human objectives get encoded into systems, who owns the resources needed to build and deploy those systems, and who has power to revise or refuse the resulting automation.
The phrase "means of prediction" gives the review its test. Data decides what becomes legible. Compute decides who can train, test, and serve systems at scale. Expertise decides who can translate objectives into models, evaluations, and deployment choices. Energy decides where infrastructure imposes cost. Treating those as neutral inputs hides political economy inside engineering procurement.
That makes the book a useful companion to Prediction Machines and Power and Prediction. Those books explain prediction as an economic input and system redesign problem. Kasy's contribution is to ask who controls the input, who chooses the objective, and who benefits when prediction becomes a governing layer.
Governance Reading
For Spiralism, the key unit is the objective file. A model does not merely predict. It predicts under a loss function, a business model, a deployment context, and a theory of whose errors matter. Hiring screens, fraud systems, recommenders, welfare triage tools, ad auctions, model agents, and pricing systems all turn objectives into social order when institutions act on outputs.
A serious review should therefore record who selected the objective, who supplied the data, what compute and cloud contracts constrain alternatives, who can inspect thresholds and evaluations, what energy or water commitments support deployment, who can challenge an outcome, and who can stop the system. Kasy's frame keeps governance from becoming ethics theater: the control question has to reach the resources, not only the model card.
This is why the book belongs near the site's pages on AI governance, AI procurement, AI system inventories, and model and system cards. The review point is practical. If an institution cannot identify the objective owner and the resource owners, it does not yet understand the system it is delegating to.
Where It Needs Care
The risk of any political-economy frame is flattening technical differences. Some systems are small, local, open, narrow, or assistive. Others are platform-scale, cloud-bound, high-stakes, and infrastructural. Data, compute, expertise, and energy are not equally decisive in every use case. The frame is strongest when it makes those differences auditable rather than when it treats AI as one block.
It also does not replace technical evaluation. A democratic objective can still be poorly measured. A public system can still fail under distribution shift. A transparency report can still miss affected people. The point is not to choose politics over evaluation. It is to ask whose politics the evaluation already contains, then test whether the evidence supports the claim.
Audit Receipt
The audit-grade sentence is: Maximilian Kasy's The Means of Prediction: How AI Really Works (and Who Benefits), published by the University of Chicago Press in 2025, reframes AI as a political economy of objectives and production resources: data, compute, expertise, and energy.
The useful receipt is simpler still: before accepting an AI objective as neutral, identify who owns the means of prediction, who defined success, who absorbs error, and who can revise or refuse the system. If those names are missing, the governance file is not complete.
Sources
- University of Chicago Press, The Means of Prediction: How AI Really Works (and Who Benefits), publisher listing, author, November 2025 publication metadata, ISBN 9780226839530, 224-page listing, description, and table of contents, reviewed June 25, 2026.
- Amazon, The Means of Prediction by Maximilian Kasy.
- Institute for New Economic Thinking at the Oxford Martin School, The Means of Prediction - How AI Really Works (and Who Benefits), author-tagged publication page and citation to the University of Chicago Press title, reviewed June 25, 2026.
- Related internal context: AI Evaluations, AI Governance, AI Procurement, AI System Inventory, Prediction Machines review, and Power and Prediction review.
Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.