The Irrational Decision—A Book Review

It’s the 5th annual April Cools! Here are my previous April Cools articles This year it’s a book review of Ben Recht’s book, The Irraional Decision: How We Gave Computers the Power to Choose For us, released Mar 10, 2026.

The publishing industry has a stupid name for the subcategory of non-fiction book where a domain expert weaves factual evidence together into a story to try to convince you of their thesis: a “big idea” book.

Generally speaking, I like books with opinions. While I didn’t enjoy the canonical examples of big idea books, like Freakonomics or Thinking Fast and Slow, I generally want nonfiction authors to express a concrete opinion about their subject. Even if I don’t agree, the expression of an opinion sharpens my attention by pushing me to think about why I agree or disagree with the author. It’s a dialogue. It’s invigorating.

So it’s not entirely a bad thing to say that my opinion of Ben Recht’s The Irrational Decision is mixed.

Recht’s core argument, which I largely agree with, is that the various incarnations of “rationalism”—which Recht paints as the root of regulatory hurdles in medicine around null hypothesis testing, public policies based on flawed behavioral economics, and empty maxims like “Trust Science”—are often overplayed and sometimes harmful.

Instead, Recht argues, we should carefully understand what classes of problems are amenable to a “mathematically rational” solution. Recht gives a few versions of the same recipe for how to tell: the objective for success must be unambiguous, measurable, and stable over time; relevant stakeholders have to agree that a mathematically rational solution suffices for practical concerns (even if it doesn’t capture everything); and you need to be able to express it in terms of optimization.¹ Problems that don’t fit in this class should be driven by moral values and require the kind of fuzzy expertise that hasn’t (perhaps, yet?) become amenable to mathematical analysis. Most importantly of all, we should collectively retain the power to change the rules as we find that they are no longer serving us.

I’m careful here to avoid mentioning computers, because this is the core of my main criticism of the book: in my view, all of this is basically unrelated to computing. Recht’s enemies are almost entirely policy decisions that at best incidentally involve computers. While he paints a historical picture of people who, enamored by computing, project utopian ideas onto what computers can do for society, those people are basically disjoint from the behavioral economists, medical boards, and business barons that use rationalism mainly as rhetoric to justify particular agendas. Or at least, if there is a direct connection, no such evidence was presented in the book.

The book is split into seven chapters, the first six of which primarily focus on the historical origins of various ideas like machine learning, linear programming, game theory, behavioral economics, and clinical trials with null hypothesis testing. I’d say this aspect was my favorite part of the book. Though I was familiar with many of the origins, such as Dantzig’s work on linear programming, there were new details about these stories I hadn’t heard of. For Dantzig, there was a parable about a team of human “computers” who worked for 120 person-days to manually apply the simplex algorithm to solve a Stigler-style diet problem, thus demonstrating the value of his simplex method and providing a baseline of cost to advocate for increased funding toward building digital computers.

Another good example was the origin of a technique called simulated annealing, which I had often heard of but never understood where it was actually useful. In Recht’s telling, it was core optimization technique of Alberto Sangiovanni-Vincentelli that, via the founding of Solomon Design Automation (SDA) in 1983, began a new era of microchip design. That story has inspired me to go look for uses of simulated annealing in modern chip optimization methods. From a cursory look, it seems to be the “classical” method for a problem called place and route, and concrete implementations can be found in software like Yosys’s nextpnr.²

But when he turns to behavioral economics and analysis of clinical trials, linear programming and simulated annealing are nowhere to be seen. Machine learning is not even present. Don’t take this to mean the topics aren’t interesting, or the writing is bad. I enjoyed Recht’s analysis of early vaccine trials, mammogram screening studies, and other clinical issues. The discussion of behavioral economics, if trite to me personally, was nevertheless cogent. These are important issues and Recht makes good points across the board.

But for a book ostensibly about society absconding decision-making to computers with ever-growing capabilities, computing is largely a rhetorical proxy for any sort of formal analysis or rule making writ large. And remember, I’m obsessed with the nitty-gritty details of how math succeeds or fails on the front lines of applications. I’m always down for a criticism of math, especially the idiosyncrasies of how it materializes in the constraints of a specific software system. But all of Recht’s examples of failures could have been done just as well on pencil and paper. Except for a brief discussion of self-driving cars, truly contemporary examples of automated decision making with computers are conspicuously absent.³ Even for the example of the National Residency Matching Program, which Recht closes Chapter 3 briefly lauding as a great success, computers were not involved until 1997, nearly 50 years after its inception. Before that, the deferred acceptance algorithm was performed via one 7-hour day of hectic phone calls. And still, there is so much more that can be said about the role of computing in the NRMP beyond the game theory! (I have two chapters about it in my upcoming book, Practical Math for Programmers).

But shining through my layers of mild disappointment, Recht embraces one more truly intriguing idea: that there are more ways to produce scientific knowledge than mainstream scientific practice acknowledges. Or to be even more brash, that science is not the only way to know something. While Recht focuses on null hypothesis testing, his narrative extends beyond science entirely. Scientific thought is basically incapable of understanding how an expert makes decisions in high-stress environments, not to mention how to formalize and automate that mechanism. Indeed, the computer here is a foil for humanity and human thought. Insofar as a computer can beat humans at chess, Recht clarifies, its mechanism for doing so is unrelated to how humans think about chess. Trying to model a computer after a human chess player’s expertise failed to produce a champion algorithm. Tree-search, self-play, value estimation via machine learning, and a lot of compute power was what worked.

Unfortunately, Recht gets lost in his musings about board games, leaving this far more interesting question open: what exactly are these different ways of knowing? What examples exist in the history of these fields he derides as having been entrapped by hypothesis testing? What examples exist outside science altogether? I suspect Recht would not appreciate the association, but in my view this mirrors similar questions (with similarly unclear answers) in books like Robin Wall Kimmerer’s Braiding Sweetgrass and Jenny Odell’s How to do Nothing, whose narratives touch on this question far away from the bureaucracy of clinical trials.

I went into this book as a frequent reader of Ben Recht’s newsletter argmin. Among the other good series on that blog, Recht wrote at length about the messy and pre-statistical-trials discovery of vitamins. He closes the first article with,

I wrote this up about four years ago, but I’ve never figured out how to publish it. Initially, I thought it would be a core part of my book on automated decision making, The Irrational Decision […], but the vitamin story ended up being too much of a prequel. The discovery of vitamins happened two world wars before the development of the computer. It happened a decade before the formalization of statistics. That it’s one of the last major discoveries of the pre-data scientific age is why I find the story so fascinating.

His initial impulse was right. This parable serves as a far better foil than the chosen side quest into board games, the pining over Shannon’s information theory, or the history of optimizing microchips. I encourage everyone to go read the blog series. I hope Recht dives further into this question in his future writing.

In his seventh and final chapter, Recht summarizes his main arguments, again primarily about policymaking and rationalism rather than computing. (Paraphrasing his points:) Society shouldn’t treat null hypothesis testing as the only way to produce or communicate scientific knowledge. Situations in which the objective or metric is subjective should not be shoehorned into mathematical models that are then treated as canon. Strict rules cannot adapt to change without explicit intervention. We shouldn’t make science writ large the sole arbiter of political decisions. People detest feeling like robots or cogs in a bureaucratic machine. Just because something can be done doesn’t mean it should be done.

In making these points, what are the examples that Recht derides? Pundits and charlatans like Steven Pinker who popularize behavioral economics. Utilitarians and effective altruists-cum-fraudsters like Sam Bankman-Fried. COVID-19 school closure policies and the “nudge politics” of Richard Thaler. Medical risk assessments, certain cancer screening practices, and inflexibility of doctors to treat patients as individuals. Mandated employee training courses and safety checklists (though only sometimes these are bad?). Some people (not specified who) don’t get enough freedom to “play” in a rigid system of rules (not clear what system).

Seeing a list of enemies like these was the nail in the coffin: computing is a sideshow. I can only guess that Recht’s focus on the academics of the 1960s and ’70s who pined for a computational utopia, which never came close to materializing, is to entice a particular target audience. Maybe that target audience still falls for the shallow rhetoric of rationalism that adorns our uniquely modern stupidities. Maybe they didn’t bother to read the texts and emails of Sam Bankman-Fried that revealed him as a selfish, manipulative prick who used effective altruism as a facade to justify his greed. Or maybe they ignored Elon Musk’s self-inflicted unraveling of his own genius myth, leaving extreme libertarianism and wealth as his only moral tenets. Maybe they forgot how many times AI barons told us AGI was 6 months away. Or maybe they missed when the tech companies that purported to “do no evil” and make only the best data-driven decisions instead optimized themselves into favoring slop over happy users, all the while regularly kissing the ring of our corrupt president to try to curry favor.

Maybe Recht already knows a more obvious rot can be found in the last 50 years of questionable economics, corrupted politics, and growth-at-any-cost business models. That aggressive regulation of the medical industry is a net good in the face of corporations that will cut all possible corners to increase profit margins. And maybe the Berkeley students he interacts with would be deaf to such arguments because they’re ramping up for internships at Jane Street, where identifying arbitrage opportunities is a holy virtue.

Perhaps this target audience needs a trusted voice. Someone who can demonstrate, over the course of two hundred engaging pages, that they really do understand the mathematics and computer science we love so much. And with that trust established, they will listen to a sober and relatively gentle criticism of that particular strain of rationalism that permates Silicon Valley. After all, rationalism is not the enemy in Recht’s book, just the people who apply it poorly.