Book Review of: B. Recht (2026). The Irrational Decision. How we gave Computers the Power to Choose for Us, Princeton University Press, Princeton & Oxford, x + 270 pp.

Voor een Nederlandse vertaling klik hier.

Dr G.J. van Bussel

The author

Benjamin Recht is Professor in the Department of Electrical Engineering and Computer Sciences at the University of California, Berkeley. He has received several prizes and awards recognizing the quality and impact of his research and has served on the editorial boards of the Journal of Machine Learning Research and Mathematical Programming. He is co‑founder of the Conference on Learning for Decision and Control. Recht’s research group investigates how to design machine learning systems that remain robust when interacting with dynamic and uncertain environments. He develops reliable benchmarks and baselines for the assessment of performance of machine learning technology. He has demonstrated, for example, that standard machine learning theory often mischaracterizes the behaviour of deep neural networks and that a number of results in personalized medicine overstate their practical utility. His research is strengthened by close collaborations with scholars in applied domains such as computational imaging and robotics. ¹

The key thesis

The key thesis is introduced and presented in the first chapter. Recht prefers the term ‘mathematical rationality’ to describe the narrow, statistical conception of reason that inspired the desire to build computers, shaped how they would eventually operate, and constrained the kinds of problems they would be considered well suited to solve. The Irrational Decision traces how, in the 1940s, mathematicians converged on a distinctive definition of rationality in which every decision is cast as a statistical problem of risk. This notion of rationality is heavily mathematical, and it has proved attractive to hyper‑online rationalists, public intellectuals, and billionaire technologists who (quite irrational) take their own success as evidence of uniquely clear thinking rather than luck or structural advantage.

Within this worldview, rationality means reducing every situation to the maximization of an objective, such that the value of an outcome is determined entirely by how well it scores on a pre‑defined metric. Intuition, experience, and judgment are displaced by optimization and statistical prediction. Recht argues that the core algorithms developed in the past underpin the automated decisions that organize much of contemporary life, from supply‑chain management and airline scheduling to the placement of advertisements in social media feeds.

For the most part, The Irrational Decision provides a historical overview of the pursuit of optimization. Its aim is not to explain how we arrived at today’s fascination with artificial intelligence, but rather to reveal what was discarded along the way and what has been lost by treating mathematical rationality as the only legitimate form of reasoning. Ultimately, Recht argues that we need less mathematical rationality. ‘We should use the tools of mathematical rationality only sometimes, in the sweet spot where they do make very good decisions — but the rest of the time, we need a human touch’ (p. 19.) That is the same conclusion as the one about statistical precision: it is better when there are clearly defined outcomes, good data, and clear reference cases that can be used for comparison (p. 187.) There are many situations in which this is not true and cannot readily be made true. Coming from a machine learning specialist, this lends his words a great deal of weight.

Recht draws a provocative thesis from this history: if ‘rational’ decisions are to be made mathematically, the world of lived experience must first be transformed into a model. Everything that is uncertain or ambiguous must be expressed in terms of probability. However, he reminds the reader insistently that the decision to create a model cannot itself be justified within the model. The decision to translate a situation into mathematics is often the truly irrational one — the point at which complexity, value and judgement risk being disregarded in favour of an apparently neat, but ultimately restrictive, calculation.

Chapter overview

Chapters 1 and 2

Having established the key thesis in Chapter 1, Chapter 2 traces the emergence of linear programming and resource-allocation models, documenting how they gradually permeated logistics, finance, public policy, and the mundane machinery of everyday administration. The chapter opens with a cautionary tale presented by Recht as a parable that encapsulates the book’s central tensions. During the Great Depression, the economist George Stigler undertook a straightforward exercise: he found the cheapest diet that would meet all the nutritional requirements of an American adult. Working entirely by hand, he calculated the annual cost to be $39.93. The resulting diet was so austere that it violated every principle of palatability, variety, and human dignity. Such a diet went against everything common sense would dictate. In 1947, Jack Laderman and his colleagues at the Mathematical Tables Project used George Dantzig’s algorithm (‘the simplex method’) to calculate the optimal solution again. This resulted in a cost of just twenty-four cents less than Stigler’s estimate but required approximately 120 person-days of desk calculator labour. The mathematics had been refined to an unprecedented level of precision, but the diet remained equally unpalatable and inhumane. Recht returns to this story repeatedly, emphasizing that mistaking mathematical optimality for genuine desirability invites consequences that are material — affecting real bodies and lives.

According to Recht, Laderman’s experiment served an unintended purpose: it demonstrated the economic case for digital computers. If enormous human labour could be justified to optimize a trivial problem, the incentive to automate such calculations was overwhelming. Thus, from their inception, the machines built to support human judgement were directed towards replacing it, a trajectory that would reshape not only computation, but also the very nature of decision-making itself. He acknowledges the successes of optimization: the areas in which it delivers remarkable efficiencies, genuine insights, and real improvements. He highlights that optimization performs impressively when the underlying problem has been carefully specified, and when the objectives and constraints accurately reflect what is at stake. However, it can perform disastrously when the specification incorporates unexamined value judgements, treating them as neutral technical choices rather than the contested ethical and political decisions they are, which demand democratic deliberation rather than algorithmic resolution.

According to Recht, the absurdity of the diet problem provides a lens through which to examine our contemporary moment: an age in which the tools of mathematical rationality have become so powerful and ubiquitous that we risk forgetting the difference between optimizing a system and serving human ends.

Chapters 3 and 4

These chapters illustrate Recht’s broader thesis that mathematical rationality has ‘sweet spots’, delivering remarkable results, but becoming misleading when applied beyond these boundaries. Although game theory and randomized clinical trials (RCTs) were conceived as universal frameworks for rational decision-making, they proved most valuable when used as tools for structuring deliberation rather than as substitutes for human judgement.

Chapter 3 traces the development of game theory from von Neumann and Morgenstern’s work in 1944 ² through its applications in military strategy and, ultimately, its unexpected success in computer gameplay. Based on a meticulous analysis of the complexity barrier, Arthur Samuel’s checkers programme, military applications, and the Prisoner’s Dilemma, the core argument is that, although game theory was conceived as a mathematical framework for understanding human economic behaviour, it has proven to be remarkably poor at predicting how people actually act. Paradoxically, game theory found its greatest success not in modelling human decision-making, but in enabling computers to master games such as chess, checkers and poker, as Recht illustrates with examples of Deep Blue, AlphaGo and Poker. The value of game theory lies not in its ability to describe human behaviour, but in its capacity to provide a framework for designing fair and transparent decision-making systems, such as the National Resident Matching Program. Game-theoretic mechanisms can resolve complex allocation problems, even when the participants themselves are not mathematically rational.

Chapter 4 examines the rise of RCTs as the ‘gold standard’ for evaluating medical interventions, exploring both their power and their limitations. Recht discusses statistical case studies, the streptomycin trial, the limits of statistics (it is effective when ‘things are highly variable but potentially predictable’ (p. 112)), vaccine trials, cancer screening, and the limits of replication. Recht’s conclusion is that RCTs excel at answering policy questions but struggle with generalization, rare outcomes, and interventions whose effects depend heavily on context. They are invaluable for reducing bias and enabling accountable decision-making, but their power lies in providing clear rules for hard decisions — not in delivering definitive answers. As with game theory, the value of RCTs is procedural: they create a framework for evidence-informed policy, but the evidence itself remains ambiguous.

Chapters 5 and 6

At this point, the book has made its historical argument overly clear. In Chapter 5, Recht argues that statistical pattern recognition and modern machine learning are not products of deep theoretical breakthroughs, but rather the result of a sustained faith in predictability, iterative engineering, and competitive benchmarking. Machine learning excels in a narrow niche where prediction is believed possible but cannot be explicitly coded, yet it lacks rigorous theoretical grounding and should not be conflated with genuine understanding or intelligence. Recht argues that machine learning is fundamentally ‘what we do when we don’t understand. When we do understand, we can just write code’ (p. 179.) In the end, machine learning is a powerful but narrowly applicable tool built on faith in predictability, competitive benchmarking, and gradient-based optimization.

In Chapter 6, Recht makes clear that human decision-making and algorithmic decision-making are not simply in competition, but that they excel in fundamentally different domains. His conclusion is that mathematical rationality wrongly assumes that all decision problems could be solved with mathematical formulas. In reality, humans and machines have complementary strengths: algorithms dominate when problems are well-specified and evaluated by averages; humans dominate when contexts are ambiguous, metrics are unclear, and expertise is contextual. Recognizing this boundary is essential to deploying automation wisely without eroding the irreplaceable value of human judgment. The people who most loudly advocate for mathematical rationality are themselves making fundamentally intuitive choices about which values to optimize for, which metrics to trust, and which human outcomes count as success.

Both chapters reinforce Recht’s broader thesis: computational tools are powerful but bounded, and wisdom lies in knowing where their sweet spots end and human judgment must begin.

Chapter 7

This chapter synthesizes the preceding six chapters by returning to the central question posed in Chapter 1: Should mathematical rationality trump our values? Recht’s answer remains ‘No’, but with nuance. The four pillars (optimization, game theory, randomized trials, machine learning) have delivered remarkable advances. The danger lies not in the tools themselves, but in mistaking their narrow applicability for universal wisdom. The path forward requires humility about what computation can achieve, clarity about where human judgment is irreplaceable, and courage to make value-laden choices even when uncertainty cannot be eliminated.

Recht argues that the most productive path forward in the tension between rule-based computational systems and human judgment lies not in choosing one over the other, but in recognizing their complementary strengths and designing interfaces that allow humans and machines to collaborate meaningfully. He critiques the assumption that humans ought to think like computers — that mathematical rationality is an end in itself. This mindset has produced harmful outcomes, like the subprime mortgage crisis, justified by ‘rational’ financial models. The chapter’s title is deliberate: ‘cyborg’ suggests integration, not replacement. The goal is not to make humans more like computers or computers more like humans, but to design systems where:

• Computers handle routine, rule-based tasks at scale
• Humans provide values, context, creativity, and moral judgment
• Interfaces allow meaningful collaboration, with humans retaining ultimate agency over value-laden choices

Recht’s final message is cautiously optimistic: we cannot compute our way to utopia, but we can be honest about what we are doing, frank about how our work impacts people, and committed to participatory processes that keep human values at the centre. The challenge of the post-computer age is not to abandon mathematical rationality, but to balance rules and play—to build systems that serve human ends rather than demanding that humans serve the logic of the machine.

Strengths

Recht has written a book that is genuinely accessible to general readers without sacrificing precision for practitioners. As a participant observer who has developed influential techniques in modern machine learning and optimization, Recht is well-placed to critique them. He leverages his professional connections extensively, speaking with colleagues to gather contemporary perspectives on the historical lineage he traces. This gives the book an authority that an outsider could not replicate. When someone who has spent their career making optimization work says that it has structural limits, the claim carries a different kind of weight. The Irrational Decision functions as an intellectual history that allows the reader to understand and empathize with the flawed logics of its protagonists even as it demonstrates the consequences of their decisions.

The book is not a polemic against mathematical rationality in all its forms. It is a precise strike against its dominance in contexts where it does not belong. Recht is careful to acknowledge where mathematical rationality genuinely works — accelerating computation, regulating pharmaceuticals, enabling electronic commerce — while arguing that it fails badly outside those sweet spots. Among the book’s most useful contributions is a clear set of conditions under which mathematical optimization is genuinely appropriate: the objective must be unambiguous, measurable, and stable over time; relevant stakeholders must agree that a mathematically optimal solution is sufficient for practical purposes; and the problem must be expressible in optimization terms without ambiguity. Problems that fail these tests require moral reasoning, contextual judgment, and the kind of experiential expertise that hasn’t yet — and may never — become amenable to formal analysis.

Recht’s prose is at its most compelling when he dismantles the idea that mathematical rationality can serve as a viable blueprint for social governance. He exposes the hubris of technocratic projects that attempt to reduce life’s inherent complexity to solvable equations, showing how rigid models inevitably break down when confronted with reality’s unpredictability. Instead of advocating a retreat from technology, he presents a nuanced argument for hybrid ‘cyborg’ decision-making systems that are designed to enhance human intuition and contextual reasoning rather than replace them with automated pattern recognition. His writing is clear, authoritative, easy to follow, and full of quotable phrases about the folly of giving optimizers and computers the power to govern our lives.

Weaknesses

While Recht’s observations cut deeply, they concentrate mainly on the inappropriate application of mathematical rationality, the transgression of knowledge boundaries, insufficient attention to technical constraints, and the acknowledgment of human values. These represent significant concerns, but the analysis overlooks structural forces, technocratic belief systems, and how these dynamics propel the very issues Recht highlights. These challenges extend beyond questions of knowledge and methodology. They are political, economic, and social in nature. They concern how mathematical rationality and the technological infrastructures built upon it have been commandeered by concentrated corporate interests to exercise social influence and accumulate vast wealth in service of reshaping society according to their vision.

I found the mismatch between the subtitle of the book, How We Gave Computers the Power to Choose for Us’, and the fact that much of the book’s argument is largely unrelated to computing quite remarkable. Recht identifies behavioural economists, medical boards, policy entrepreneurs and rationalizing billionaires as using mathematical language primarily as rhetoric to legitimize agendas formed on other grounds. These groups are largely disjoint from the engineers and computer scientists who built the optimization machinery that Recht knows best. The expectation that I had based on the subtitle was not consistently fulfilled by the text. Promise and reality do not match.

What is equally striking is what the book leaves out: contemporary artificial intelligence. For a work appearing in 2026, at the zenith of large language models and algorithmic organizational decision-making, the scarcity of current examples is remarkable. Self-driving cars receive attention, yet recommendation engines, predictive policing algorithms, and automated credit-scoring systems go unexamined. These are precisely the domains where Recht’s critique would resonate most powerfully. Their absence represents, at minimum, a significant missed opportunity.

The book’s argumentative force could have been considerably greater. Recht’s passing reference to the history of vitamin discovery (p. 21) — arriving before statistical methods and computational tools — hints at a more compelling historical counterpoint than chess could provide. A fuller exploration of how humans uncovered objective nutritional truths long before developing the statistical machinery to formalize those discoveries would have bolstered Recht’s central claims far more effectively than game theory examples ever could. Such a narrative would demonstrate that human insight and empirical observation can yield genuine knowledge independent of mathematical formalization, strengthening the case for the irreplaceable value of human judgment in ways that abstract game-theoretic models cannot.

Data-Driven Decisions and the Irrepressible Role of Intuition.

Benjamin Recht’s The Irrational Decision strikes a profound chord by illuminating the indispensable role of intuition in human judgment, even when this phenomenon remains largely unnamed. At its core, the book explores a fundamental tension that cognitive scientists have long examined: the conflict between deliberate, rule-based, explicit calculation on one side, and tacit, experiential, pattern-recognition-based judgment on the other.

The Stigler diet problem exemplifies this tension perfectly. The linear program is solved; the optimal solution is calculated; the mathematical proof is complete. Yet when a human examines the output and declares, ‘no one will eat this,’ that rejection is not itself a mathematical act. It is a judgment grounded in embodied knowledge, social understanding, cultural context, and practical wisdom. The model was mathematically correct. The human intuition that rejected it was also correct—and it operated on an entirely different register. Neither could substitute for the other.

This dynamic recurs throughout the book across different domains. In clinical trials, the formal apparatus of statistical inference proves fundamentally inadequate for capturing how experts make decisions under pressure, let alone how to automate or replicate such processes. A physician interpreting a p-value does so through the lens of years of clinical experience, intimate knowledge of the individual patient, and pattern recognition honed from thousands of prior cases. The number does not speak for itself; the doctor must translate it, and that translation constitutes the genuine challenge.

In behavioural economics, Richard Thaler demonstrated that humans systematically deviate from the mathematical rational actor model. ³ What Recht illuminates is the profound irony: those who most vociferously proclaim the necessity of mathematical rationality are themselves making fundamentally intuitive choices about which values to optimize, which metrics to trust, and which human outcomes constitute success. Their mathematical apparatus rests upon a foundation of prior judgments that the mathematics itself cannot validate.

This represents the book’s deepest insight, and it connects to a substantial body of research that Recht gestures toward without directly citing. Gary Klein’s work on naturalistic decision-making demonstrates that expert practitioners — from firefighters to intensive care physicians — make rapid, effective decisions not by performing mental expected-utility calculations, but by recognizing situations as instances of previously encountered patterns and mentally simulating courses of action forward. ⁴ Daniel Kahneman’s framework of System 1 and System 2 thinking captures a related insight: fast, associative, intuitive cognition is not the enemy of sound decision-making but rather its precondition. ⁵ According to Matthew Lieberman, when making choices and decisions, our neural systems for reasoning and logic are deeply intertwined with our innate drive for social understanding, empathy, and collaboration. Decisions based solely on mathematical calculation are simply impossible. ⁶

What Recht contributes is a structural critique that extends beyond cognitive science: the danger lies not in humans using intuition, but in institutions that pretend intuition does not exist, then construct systems that compel decision-makers to simulate mathematical procedures while actually performing something far more human beneath the surface. The physician checks the protocol box while knowing from experience that this particular patient requires something different. The analyst presents the model’s output while having privately concluded it does not fit the case. The cognitive labour of intuition becomes hidden, delegitimized, and consequently unexamined, which means it also goes uncorrected. ⁷

The Irrational Decision reveals why only humans can resolve fundamentally political or value-laden questions and proposes a more expansive approach to decision-making: one appropriately supported by computational tools yet firmly grounded in human intuition, moral judgment, and institutional accountability. It requires accepting that mathematics is genuinely powerful within specific, well-defined domains while insisting that domain-setting, value-weighting, and final interpretation remain irreducibly human acts—and should be recognized as such.

Conclusion

The Irrational Decision is a valuable, timely, and unusually honest book about the limits of the field Recht himself has helped build. Its greatest strength is its refusal to be a polemic. For an opinionated, informed, no-nonsense account of what optimization actually does and what purposes it serves, written by someone who knows this subject from the inside, this is essential reading for anyone in data science, policy, or AI, with the caveat that its most important argument is sometimes buried under the historical scaffolding.

---

1. Online sources, retrieved 23 May 2026, from: https://people.eecs.berkeley.edu/~brecht/bio.html and https://vcresearch.berkeley.edu/faculty/benjamin-recht. ↩︎
2. J. Von Neumann, and O. Morgenstern (1944). Theory of Games and Economic Behavior, Princeton University Press, Princeton. Used in its 6^th printing: 1953. Online source, retrieved 23 May 2026, from: https://dn720001.ca.archive.org/0/items/in.ernet.dli.2015.215284/2015.215284.Theory-Of.pdf. ↩︎
3. R.H. Thaler, and C. Sunstein (2021). Nudge: Improving Decisions About Health, Wealth, and Happiness, Penguin, New York; and R.H. Thaler (2015). Misbehaving: The Making of Behavioral Economics, W.W. Norton & Company, New York. ↩︎
4. G.A. Klein (2008). ‘Naturalistic Decision Making’, Human Factors. The Journal of the Human Factors and Ergonomics Society, Vol. 50, No. 3, 456–460, and G.A. Klein, J.M. Orasanu, R. Calderwood, and C. Zsambok (1993). Decision Making in Action: Models and Methods, Ablex, Norwood (NJ). ↩︎
5. D. Kahneman (2011). Thinking, Fast and Slow, Farrar, Straus and Giroux, New York. ↩︎
6. M.D. Lieberman (2013). Social. Why our Brains are Wired to Connect, Crown, New York. ↩︎
7. Which shows itself also from other perspectives. Over-organizing organizational business processes tends to create spaces of contestation, in which employees negotiate about how to decide and what decisions to make. Those spaces deliberately ignore procedures and decisions are most often ‘retrospectively inscribed’ in systems. G.J. van Bussel (2020). A Sound of Silence. Organizational Behaviour and Enterprise Information Management. Papers on Information and Archival Studies, I, Van Bussel Document Services, Helmond, pp. 89-91. ↩︎