The Means of Prediction (Maximilian Kasy)

Martin Swartz
51 minutes ago
17 min read

In this Book Essential

INTRODUCTION

Every wave of AI hype arrives with the same story: machines are about to surpass us, and their rise is as inevitable as the weather. Maximilian Kasy, a professor of economics at Oxford who teaches machine learning theory to graduate students, thinks that story is not just wrong but convenient. In The Means of Prediction: How AI Really Works (and Who Benefits) (University of Chicago Press, 2025), he argues that the real question about AI was never technical. It is a question of power: who gets to pick the objectives that AI systems optimize, and who captures the benefits.

The book is built on a single, sharp reframing. AI is automated decision-making that maximizes a measurable objective, and building it requires four resources Kasy calls the means of prediction: data, computational infrastructure, technical expertise, and energy. Control those resources and you control what AI does. In a capitalist economy, that control sits mostly with the owners of capital, which is why feed algorithms amplify outrage, hiring algorithms screen out future parents, and warehouse algorithms push workers to injury. These are not failures of engineering. They are systems doing exactly what they were built to do.

Kasy's target reader is anyone who has felt that AI debates are too technical to join. He spends the first half of the book proving otherwise, walking through supervised learning, overfitting, deep learning, and exploration versus exploitation in plain language, with Babylonian omens, Odysseus, and the Mechanical Turk as guides. The second half turns those foundations into political economy: how AI concentrates wealth, why privacy law cannot stop it, what automation really does to wages, and why he believes democratic control of AI objectives is the only durable answer.

This Book Essential is for you if you want to understand AI well enough to argue about it: what machine learning actually computes, why more data changes what algorithms can do, where the technology's real limits sit, and who is currently deciding what all of it is for.

U365'S VALUE PROPOSITION

WHO THIS IS FOR

Professionals and managers who deploy AI tools at work and want to understand what the systems they buy are actually optimizing.
Policy-minded readers, journalists, and civic technologists who need a rigorous but accessible framework for AI regulation debates.
Students of economics, political science, and computer science who want one book connecting the technical and political sides of AI.
Founders and product leaders deciding what objectives their algorithms should pursue, and who should have a say in that choice.
Educators designing AI literacy programs for non-engineers.
Citizens who suspect that "AI is too complicated for public debate" is a story that serves someone, and want evidence.

KEY TENSIONS

Optimization versus conflict: Machine learning frames every problem as the maximization of a single objective. Society's problems are conflicts between people with different objectives. Kasy argues that treating social questions as engineering problems erases the question of who wins and who loses.

Fate versus choice: The dominant AI story says progress is inevitable and beyond control. Kasy insists every deployment is a human decision, repeated over and over, and that inevitability talk preserves the status quo for those it benefits.

Alignment with whom: The standard framing asks how to align machine objectives with human objectives. Kasy flips it: the hard conflicts are between different humans, and "whose values" is a question of control rights, not engineering.

Privacy versus learning: Differential privacy protects the individual in a dataset. But machine learning is about patterns across people, so even perfectly private data collection can produce predictions that harm you through other people's data. Individual consent cannot govern collective consequences.

Automation versus augmentation: The same technology can raise workers' productivity and wages, or eliminate their jobs. Which one happens depends on who chooses the objectives, not on the technology itself.

Fairness versus welfare: Standard algorithmic fairness defines bias as a deviation from profit maximization. A welfare lens asks instead who gains and who loses, and whether that distribution is acceptable. The two frameworks can point in opposite directions.

WHY IT MATTERS NOW

AI is moving from feeds and recommendations into consequential decisions: hiring, credit, insurance pricing, medical triage, education, and warfare. Kasy's framework lands differently once you notice that every one of these systems encodes a choice of objective that someone made, usually without the affected people in the room.

The book also speaks to the current governance moment. The EU AI Act, copyright litigation over training data, antitrust action against platform monopolies, and debates about AI's energy footprint are all, in Kasy's terms, fights over the means of prediction. Reading them through his lens turns scattered policy stories into one coherent contest.

Finally, the book is a direct rebuttal to AI fatalism. If AI's direction were fixed by technology, public deliberation would be theater. If it is fixed by ownership and control, then collective action, labor law, tax policy, and democratic innovation become the levers that matter. Kasy wrote the book to make that second view arguable by anyone willing to learn the mechanics.

OVERVIEW

Kasy organizes the book in five parts. Part I dismantles the "humans versus machines" story that Hollywood and Silicon Valley tell, replacing it with a conflict between people over objectives. Part II teaches how AI actually works: decision problems, supervised learning, the overfitting and underfitting trade-off, deep learning, and the exploration versus exploitation balance, ending with a summary chapter of key ideas.

Part III, the book's core, builds the political economy. It starts with social welfare as the yardstick for evaluating AI's consequences, then names the four means of prediction, surveys who might realistically change AI's direction (workers, consumers, journalists, judges, regulators), and dissects the ideological stories that make change seem impossible.

Part IV applies the framework to five regulatory domains: value alignment, privacy, automation, fairness, and explainability. Each chapter lands on the same conclusion: engineering fixes alone cannot work, because the underlying issues are conflicts of interest that require democratic resolution. Part V closes with the ancient questions AI renews, and sketches institutional forms democratic control could take, from sortition to liquid democracy.

The distinctive move of the book is to treat AI as economics rather than as science fiction. Kasy brings contract theory (incentive design, multitasking), externalities, Pigou taxes, marginal productivity, and welfare weights to bear on algorithms. The result is a book that respects the reader's intelligence on both the math and the politics, and refuses the premise that either is too complicated for public debate.

KEY IDEAS

MEANS OF PREDICTION

Data, computational infrastructure, technical expertise, and energy are the four resources needed to build AI. Kasy names them the means of prediction, echoing Marx's means of production, because ownership of them determines control over AI's objectives. The chapter on enclosure walks through how data went from commons and personal life to corporate assets, how compute concentrates in a few cloud providers, and why energy is unlikely to be the binding constraint.

AI AS OPTIMIZATION

Kasy defines AI as the construction of systems for automated decision-making to maximize a measurable reward. Four ingredients define every decision problem: the action space, the reward, prior knowledge, and data. Somebody must choose the reward, and that choice, not the algorithm, is where the politics lives.

SUPERVISED LEARNING AND PREDICTION

Most modern AI is prediction: features X go in, a label Y comes out. Facial recognition, language models, ad targeting, hiring screens, and bombing target selection are all the same mathematical shape. The chapter grounds this in the ancient dream of divination, from Babylonian omens to crystal balls, then shows how training on large datasets replaced them.

OVERFITTING AND UNDERFITTING

Every learner must sail between Scylla and Charybdis: overfitting (developing superstitions from random patterns) and underfitting (refusing to learn from evidence). Model complexity is the lever, and cross-validation, testing on held-out data, is the compass. More data shifts the trade-off toward flexibility, which is why the big data era favored complex neural nets.

DEEP LEARNING AND SCALING

Neural nets build complicated functions from simple ones, layer by layer, learning by gradient descent. Kasy demystifies the artificial brain metaphor: modern deep learning is craft, not biology. Its success came from scaling data and compute (Moore's law, the internet, GPUs), not from conceptual breakthroughs. Transformers power language models; diffusion models power image generation; both are self-supervised tricks that turn unlabeled data into training signal.

EXPLORATION VERSUS EXPLOITATION

When actions change what you observe (multi-armed bandits, drug trials, ad placement), algorithms must balance trying new options against using what works. Reinforcement learning adds planning, which is how AlphaGo mastered go through self-play. The same algorithms that recommend your feed also decide clinical trials, which is why the objective choice matters everywhere.

IDENTIFICATION AND ALGORITHMIC EQUIVALENCE

With enough data, different algorithms converge on the same predictions. So public debate can ignore the algorithm and focus on the decision problem: what is being predicted, for whom, and to whose benefit. This is the book's democratic escape hatch: you do not need to understand transformers to govern what they optimize.

SOCIAL WELFARE AS THE YARDSTICK

A society is good if it is good for the people in it. That premise forces three questions: whose welfare counts, how to measure it, and how to weigh one person's welfare against another's. Welfare weights make the trade-offs explicit, where machine learning's single objective hides them. Kasy contrasts utility-based welfare with Rawls's primary goods and Sen's capabilities.

AGENTS OF CHANGE

Change will not come from asking engineers to be nice: corporations are profit-maximizing by law and by competition. Agents of change need bargaining power (strikes, boycotts, bad press, litigation, regulation), values, and interests. Kasy surveys workers and unions, consumers, journalists and media, judges, policymakers, and the law's levers: intellectual property, privacy, antitrust, interoperability standards, labor law.

IDEOLOGICAL OBFUSCATION

Three ideological moves block change: claiming a general interest where there is conflict, denying alternatives, and reframing social choices as technical issues. The humans-versus-machines story performs all three: it erases conflicts of interest, makes AI's direction seem inevitable, and implies only experts can decide.

VALUE ALIGNMENT AS INCENTIVE DESIGN

The mis-specified objective problem (the paper-clip maximizer) is the classic multitasking problem of contract theory: you cannot reward what you cannot measure, and high-powered incentives distort. Inverse reinforcement learning cannot fix it, and raises the deeper question: alignment with whom? Control rights, not cleverer rewards, decide whose values count.

PRIVACY EXTERNALITIES

Differential privacy guarantees your inclusion in a dataset is undetectable, so you have no personal incentive to withhold data. But machine learning extracts patterns across people: even if you never share your health data, others' data lets insurers predict your risks. Because externalities are the whole point of learning, only collective governance, Pigou taxes, data trusts, regulation, can set the right level of data collection.

AUTOMATION AND MARGINAL PRODUCTIVITY

Wages track marginal productivity, not average productivity, so technology can raise output per worker while cutting some workers' pay. History (the Luddites, office computing) shows both automation and augmentation are possible; which one happens depends on who directs the technology. Worker voice, codetermination, and the Scandinavian collective resources approach are how objectives change.

FAIRNESS VERSUS WELFARE

Becker defined discrimination as a departure from profit maximization; King defined it as the opposite of democracy and racial justice. Most algorithmic fairness work inherits Becker's lens, which means an algorithm can be perfectly fair while reproducing any level of inequality the profit-maximizing prediction implies. The welfare lens asks the question fairness avoids: who wins, who loses, and is that acceptable.

EXPLAINABILITY OF DECISION PROBLEMS

There are three things to explain: decision functions (for engineers), individual decisions (counterfactuals for those affected), and decision problems (what is being maximized, over what actions). For democratic governance, the third matters most, and it is the one that does not require mathematics to understand.

DEMOCRATIC CONTROL

The conclusion argues democratic control of AI objectives is the only durable alignment of AI with social welfare, and sketches forms beyond electoral democracy: sortition (Landemore's open democracy), liquid democracy (with delegation), and extending democracy into the workplace. Participation without real power, Kasy warns, is bound to fail.

SUMMARY

Mind map of The Means of Prediction (Maximilian Kasy) showing chapter branches and key concepts — The Means of Prediction (Maximilian Kasy) - Mind Map

Chapter 1: The Story of Humans Versus Machines

The dominant AI narrative, from HAL 9000 to ChatGPT extinction talk, frames the future as conflict between humans and machines. Kasy reads this story as ideology: it makes AI's direction seem fated, hides who wins and loses, and serves those benefiting from the status quo. The academic variants (existential risk, value alignment, mass unemployment) share the same man-versus-machine frame.

Chapter 2: What the Old Story Misses

Technology is not fate; it is choices made repeatedly. AI is automated decision-making under optimization, and someone picks the objective. In a capitalist society, objective-setting power tracks capital: the owners of the means of prediction decide. The remedy is democratic control of objectives, and the precondition is public understanding, because AI is not actually too complicated to debate.

Chapter 3: What This Book Does

The book has two halves: how AI works (decision problems, prediction, overfitting, deep learning, exploration) and AI in society (welfare, power, regulation, ideology). The connecting thread is the choice of objectives, and the recurring conclusion that social problems with AI are conflicts of interest, not optimization errors.

Chapter 4: What Is Artificial Intelligence?

Intelligence is multidimensional: people and machines each have strong suits and weaknesses, so "more intelligent than humans" has no fixed meaning. Kasy defines AI as the construction of systems for automated decision-making to maximize a measurable reward, and walks through the four ingredients of a decision problem: actions, rewards, prior knowledge, data. Machine learning displaced expert systems because data and compute scale while human expertise does not.

Chapter 5: Supervised Learning

Prediction is the core task: features X in, label Y out. Applications range from ad clicks to job screening, bail decisions, credit scoring, university admissions, and the Lavender system used to select bombing targets in Gaza, with an accepted error rate of 10 percent false positives. With unlimited data, all algorithms converge on the same answer (identification), so the algorithm matters less than the decision problem, which is exactly what public debate should examine.

Chapter 6: Overfitting and Underfitting

Learners must steer between superstition (overfitting random patterns) and stubbornness (underfitting, refusing to update). Model complexity is the lever; Occam's razor is the principle; cross-validation on held-out data is the method, because hindsight is easy. More data reduces overfitting risk, which explains the shift to ever larger and more complex models.

Chapter 7: Deep Learning

Neural nets compose simple functions into complicated ones, trained by gradient descent, made scalable by GPUs. Deep learning is not an artificial brain: it is the craft of function-building. Self-supervised learning turned unlabeled text and images into training data, powering transformers (language) and diffusion models (images). Scraping the internet to train generative models amounts to enclosure of the data commons, a massive concentration of the key means of prediction.

Chapter 8: The Exploration/Exploitation Trade-Off

Bandit algorithms balance trying new options against using known-good ones, from pizza versus Ethiopian food to drug trials to ad placement. Reinforcement learning adds planning and state, which is how AlphaGo mastered go by self-play. It works where simulation is possible and fails where it is not: no fleet of self-driving cars can learn by crashing in the real world, and rare diseases and macroeconomics have hard data limits.

Chapter 9: Key Ideas to Remember

A recap chapter: AI is optimization; prediction needs the overfitting and underfitting balance; acting needs exploration and exploitation; planning needs delayed rewards. These mechanics matter politically because they identify what AI can and cannot do, and which resources (data, compute) control it.

Chapter 10: Social Welfare

To evaluate AI's consequences you need a definition of a good society. Kasy's premise: a society is good if it is good for the people in it. That forces three questions: who counts, how to measure welfare, how to weigh people against each other. Utility-based welfare (preferences over choices) competes with Rawls's primary goods and Sen's capabilities. Dangerous objectives (maximize employment, minimize pregnancies) show why welfare, not proxy targets, must remain the yardstick.

Chapter 11: The Means of Prediction

The book's title chapter. The four resources, data, computational infrastructure, technical expertise, energy, differ in who controls them, how far they scale, and what externalities they generate. Data is the most important and most contested; compute concentrates through cloud oligopoly; expertise is currently scarce but fungible; energy constrains climate, not AI. The enclosure of the data commons is the primitive accumulation of AI.

Chapter 12: Agents of Change

Engineers cannot save us: corporate profit maximization binds them. Effective agents need bargaining power over profits: workers (tech, click, gig, warehouse), consumers (constrained by engineered network effects, countered by interoperability rules), journalists, judges, policymakers. Levers include IP law, privacy law, antitrust, interoperability standards, and labor law. Strategic threats align objectives short-term; long-term alignment needs democratic control.

Chapter 13: Ideological Obfuscation

Ideology serves interests by claiming general interest, denying alternatives, and technologizing social choices. The humans-versus-machines story does all three, and "if we don't, China will" repeats the trick politically. Alternative stories must make conflicts of interest explicit, which is what this book attempts.

Chapter 14: Value Alignment

The paper-clip maximizer is a multitasking problem: what cannot be measured cannot be rewarded, so any single measurable objective distorts. Inverse reinforcement learning (learning objectives from human behavior) hits the deeper question, alignment with whom. Zuckerberg's feed is aligned with Zuckerberg; the question is whether it is aligned with everyone else. Short-term fixes align profits with welfare (Pigou taxes, bargaining power); long-term fixes redistribute control.

Chapter 15: Privacy

From the randomized response method to differential privacy, technical privacy is real and individual. But learning is about externalities: patterns across people. A fully privacy-compliant world can still price you out of insurance based on others' data. GDPR-style individual consent and property rights therefore cannot govern data collection; Pigou taxes, data trusts, and collective governance can.

Chapter 16: Automation

From the Luddites to Goldman Sachs' 300-million-jobs prediction, automation panic is perennial. Economics says the outcome is not fixed: marginal productivity, not average, determines wages, so growth without shared prosperity is possible and common. Recent evidence suggests generative AI helps novices most, a floor-raising effect. Worker voice (codetermination, participatory design) is how societies choose augmentation over automation.

Chapter 17: Fairness

Becker's definition of discrimination as a deviation from profit maximization shaped both economics and algorithmic fairness. Its blind spot: a perfectly profit-maximizing algorithm can reproduce arbitrary inequality. The welfare lens asks who gains and loses, with welfare weights for the worst off. Equal treatment produces unequal outcomes because starting points differ; only a welfare lens can evaluate that.

Chapter 18: Explainability

Kafka's Josef K. is the emblem of opaque decisions. Three kinds of explanation: decision functions (engineers), individual decisions (counterfactuals, adversarial examples), and decision problems (what is maximized over what options). Counterfactual explanations offer recourse but founder on the ambiguity of "small changes." For democratic governance, explaining the decision problem is the necessary and sufficient level, and it needs no mathematics.

Chapter 19: The Ancient Questions Behind AI

The recap chapter ties the mechanics to permanent questions: how to learn (overfitting, exploration), how to act (planning, incentives), what makes a good society (welfare, weights), who decides (power, control of objectives). AI does not create these questions; it re-poses them with higher stakes.

Chapter 20: Toward Democratic Control of the Means of Prediction

The conclusion argues democratic self-rule over AI objectives is the only durable solution, then works through institutional forms: sortition and deliberation (Landemore), liquid democracy with revocable delegation, extending democracy beyond the state into firms and platforms. The warning: participation without real power is a spectacle that dwindles. Democratic control needs to be fought for and built, in practice, by all of us.

IN PRACTICE

1. Ask the objective question first: When you meet any AI system, ask: what is it maximizing, over what actions, based on whose data? Kasy's framework shows this question can be answered without mathematics. Action: pick one AI tool you use weekly and write one sentence naming its likely objective and who benefits.

2. Distrust inevitability talk: "AI is coming whether we like it or not" is an ideological move that preserves the status quo. Treat every deployment as a decision someone made and could revisit. Action: when you hear inevitability framing in a meeting or article, restate the claim as a choice: who decided this, and what would changing it require?

3. Test for unmeasured objectives: The paper-clip problem is the multitasking problem: whatever is not in the reward function gets sacrificed. Before optimizing anything, list the outcomes that matter but are not measured. Action: in your next project plan, add a line naming one important outcome the current metrics do not capture.

4. Evaluate fairness by welfare, not profit: An algorithm can be unbiased by the profit-maximization definition while entrenching inequality. Ask who gains, who loses, and how much weight the losers get. Action: when reviewing an algorithmic decision system, write down the winners and losers explicitly, including people outside the user base.

5. Treat data as collective, not personal: Privacy compliance does not govern externalities: your risk comes from patterns in other people's data. Individual consent is the wrong scale for collective effects. Action: in data governance discussions, move the question from "did users consent" to "who is harmed downstream by the predictions this data enables."

6. Use cross-validation thinking in your own judgments: Test beliefs on data you have not already used to form them; that is the only honest check against overfitting your own reasoning. Action: for one strong opinion you hold, name a test it has not yet passed and what evidence would change your mind.

7. Find your channel as an agent of change: Kasy's list: workers organize, consumers boycott, journalists expose, judges rule, regulators constrain. Individuals gain power through these channels, not through the personal virtue of engineers. Action: pick the channel closest to your role and take one concrete step in it this month.

QUOTES

"Technology is not fate. Just as people make technology, people decide how it is used and what interests it serves."

"The important question, then, is who gets to pick the objectives of AI systems."

"Doing nothing serves the interests of those who benefit from keeping things as they are."

"The real conflict is not between a human and a machine but between the different members of society."

"AI is, at its core, automated decision-making using optimization. That means that AI algorithms are designed to make some measurable objective as large as possible."

"No matter who you are, don’t let anyone tell you that you are not the “type” to understand AI."

"If an algorithm selecting what you see on social media promotes outrage, thereby maximizing engagement and ad clicks, the problem is not an optimization error: Promoting outrage is good for profits from ad sales, even if it is bad for society."

"We live in a capitalist society, and in such a society the objectives of AI are typically determined by the owners of capital."

"All humans, and all machines, have their strong suits and their weaknesses."

"To summarize: Machines are being trained to behave so that they seem like humans, by using data from a platform where humans are presented as machines, but the humans behind this platform have started to substitute themselves with machines pretending to be human."

"In machine learning, there is a single objective that the algorithm must maximize. In a framework with a single objective, the worst that can happen is that we fail to optimize the objective; in such a framework, we do not have a language to talk about the conflicting objectives of different people."

"Any decision, whether human or automated, will in general lead to gains and losses for different people. AI decisions, in particular, typically increase the welfare of some and decrease the welfare of others."

"The most salient value-alignment problems are thus not between a human and a machine. They are between different humans."

"The thing I saw at Facebook over and over again was there were conflicts of interest between what was good for the public and what was good for Facebook. And Facebook, over and over again, chose to optimize for its own interests, like making more money."

"Externalities are the whole point of machine learning. Machine learning is never about individual data points; it is always about patterns across individuals."

"By making arguments about the general interest, the absence of alternatives, and the technical nature of the issues at hand, ideology has the function of preventing change and preserving the status quo."

"Collective self-governance in the form of democratic control of AI objectives thus seems ultimately the only solution that can guarantee broadly beneficial uses of AI."

"The goal is not to recover some deeper truth about the world but instead to act successfully in it."

"Democratic governance of AI needs to be fought for and put into practice. This will be a task for all of us."

AUTHORS EXPERTISE

Maximilian Kasy is a professor of economics at the University of Oxford, where he teaches machine learning theory to graduate students and coordinates the machine learning and economics group. His training spans mathematics, statistics, and economics, and much of his research concerns statistical methodology, including work on regularization and the risk of machine learning in empirical economics published in the Review of Economics and Statistics.

A second line of his research studies economic inequality and policy, including pilot job-guarantee and basic-income programs. That combination, rigorous ML theory plus distributional economics, is exactly the vantage point The Means of Prediction is written from: he can explain what a transformer computes and then trace who captures the value it creates. His publications include field experiments on adaptive job search assistance for refugees and papers on decision theory and Bayesian inference.

The Means of Prediction is his book for general readers, published by the University of Chicago Press in 2025. It deliberately teaches AI's mechanics in plain language before arguing its politics, on the premise that a democracy cannot govern what it does not understand.

RESOURCES

The Means of Prediction: How AI Really Works (and Who Benefits) by Maximilian Kasy: https://press.uchicago.edu/ucp/books/book/chicago/M/bo220760899.html

The Means of Prediction: How AI Really Works (and Who Benefits) by Maximilian Kasy: https://www.amazon.com/Means-Prediction-Really-Works-Benefits/dp/0226840410

Complementary reading: Human Compatible by Stuart Russell, the value-alignment case Kasy engages and challenges.

Complementary reading: Power and Progress by Daron Acemoglu and Simon Johnson, on the thousand-year struggle over technology and prosperity.

Complementary reading: The Atlas of AI by Kate Crawford, on AI as an extractive industry.

NEXT STEPS

Question the objective: For every algorithm that touches your life this week, ask what it maximizes and for whom. One question, asked consistently, is the whole method.

Learn the mechanics once: Read Part II of the book until prediction, overfitting, and exploration are clear. These three ideas carry the entire political argument.

Spot the ideology: When you hear "inevitable", "no alternative", or "too technical for you", name it as an obfuscation move and ask who benefits from the status quo.

Pick your lever: Choose one channel from Kasy's agent list (union, consumer action, journalism, law, policy) and take a first, small, real action in it.

Demand decision-problem explanations: In any AI procurement or review, ask for the objective, action space, and data sources in writing. This is the explainability that enables governance.

Teach one person: Explain the means of prediction to a colleague or student this month. Public deliberation scales one conversation at a time.

INSIDE

PUBLICATIONS