WASHINGTON – July. 18, 2026 – Quillette — Washington, DC
A review of The Modern Mythology of AI by Peter L. Levin.
A review of The Reverse Centaur’s Guide to Life after AI by Cory Doctorow; 240 pages; Verso (June 2026)
In late August 2010, I was a senior Obama-appointed official in the office of the Secretary of Veterans Affairs, and I received a message informing me that the secretary wanted to see me right away. These encounters never start well and usually end with someone’s head on a pike. The nerve centre of a 400,000-person agency, that office works on dozens of projects at the same time, and I had no idea which one had just blown up.
VA used to be a non-political and mostly bipartisan place. There were very few things we did, or could do, that had a “D” or an “R” attached to it. There is no better mission in government than providing health services and earned benefits to people who volunteered—and sometimes risked their lives—to serve in uniform. When I beamed myself into the secretarial suite, he was already with another official whose job I can best describe as “don’t let stupid things happen.” Both looked grim. I started doing mortgage recalculations in my head.
A couple of weeks before the secretarial summons, the president had announced that VA was going to make veterans’ health records available to them at “the push of a button.” If a veteran wanted their longitudinal health record—including all prescriptions, lab results, allergies, diagnostic images, and clinical notes—we were going to provide it, in an easy-to-read computer file. No special tools, applications, or extra permissions would be required. This was the genesis of the Blue Button program, which Craig Newmark described as “VA’s gift to the country.”
Today, most people in the United States, and anyone who gets federally provided clinical services, have access to their health records, including members of the armed forces and their families, as well as Medicaid and Medicare beneficiaries. VA was a pioneer in making all this possible. But on that day, in the secretary’s office, the concern was about how this new technology might embarrass the president and cost him votes in the next election.
The core problem was that the assistant secretary for error avoidance was convinced that “nobody thinks this is a good idea.” They had taken an informal poll of stakeholders—without the benefit of a good explanation—and all (uh, six) of the respondents had said they saw no need or purpose for access to their data, so why risk the leak of medical information? I said, “I could go to the park across the street and find six people who think trees are a bad idea. That doesn’t mean we don’t need trees.”
The secretary, of course, was in a bind. The program was about to go live, and all the announcements and press releases and promotions were locked, loaded, fuelled, and ready to launch. Together we hit upon an idea. What if we launched without fanfare? The president had already announced it, of course, but who would remember? Who would care? How would anybody find out? I lunged for the compromise. I knew that our primary stakeholders, the veteran-service organisations like American Legion and Veterans of Foreign Wars, scoured the website every day for changes anyway, and I was confident that someone would find it, whether we promoted it or not. Everybody agreed to move forward.
Within a few days, people were chattering about how they could get their records, bring them to their doctor, or pharmacist, or caregiver, and see things about themselves that they had never seen before. The scheme was a resounding success. The product-release strategy—a kind of manual “screen scrape”—worked like a charm.
In his wonderful new book The Reverse Centaur’s Guide to Life After AI, Cory Doctorow tells us that the most important fact about a technology isn’t what it does, it’s who it does it for, and who it does it to. For veterans, and later for the country, Blue Button was a blow-out win; it helps millions of people manage their care, and it harms nobody.
Today, of course, screen scrapes that mimic human reading are fully automated, and the robots hoovering up the information are not surveilling for new digital services. They’re indiscriminately looking for any kind of text, regardless of source, quality, veracity, or intent. Indeed, the entire internet is the training ground and the raw material for a new kind of application: large language models or LLMs.
Doctorow’s evaluation of LLMs has a harsh outcome. Unlike veterans, service-members, and Medicare and Medicaid enrollees, the main beneficiaries of LLMs are a very concentrated group of investors, inventors, and insiders; most of society will suffer dilatory consequences. More distressing, according to Doctorow, are the hapless employees whose jobs, their bosses think, will be replaced by the chatbots.
Many technologies have an indisputably positive impact. Wireless communications, satellite navigation, and advances in life science, transportation, fixed nitrogen, electric power, and semiconductor devices come to mind. Each of these inevitably carries its own set of liabilities too—from economic displacement to environmental degradation—but only flat-earthers would want to return to an era where we didn’t have GPS, vaccines, air travel, abundant food, and ubiquitous light switches and sockets.
LLMs, on the other hand, suffer three enormous deficits. The first is the perverse disincentive of human employment. The second is that LLMs cannot do what their hucksters say they can do. Consequently, the target customers—businesses that run call centres, for example—think they’re buying expert automation, which is a ChatGPT-class hallucination. Some businesses have already rehired the workers they fired.
The third is a more transcendental defect: AI is the first technology that people have imbued with intelligence, sentience, consciousness, autonomy, and feelings. When august publications like the Economist speak of “AI models’ values,” misperception and moral sepsis infect the bloodstream of our most reliable platforms. Unless and until policymakers, business leaders, and ordinary citizens understand what AI is and isn’t, the furnace of misplaced capital will roar and—like the housing crisis twenty years ago—consume public attention and taxpayer resources.
The fundamental element of an LLM, or any kind of generative or analytical AI, is called a perceptron, sometimes referred to as a synthetic neuron. Similar to the thinking cells of a brain, it is activated when an ensemble of its inputs are turned on.
Suppose you’re in a completely dark room, facing a wall, with your hand on a light switch in front of you. Other people you trust are in the room with you, looking out of windows you cannot see because your back is to them. They’re going to tell you if there are any other lights on at houses in the neighbourhood, even if they only flicker for a moment or two. Your job is to keep track of what they see, and to decide whether, and when, you should turn your light on too. The rule might be to light up your room if three of your four friends report a light in another house within a few seconds of each other.
We’ve just described a very primitive neuron. Live models are infinitely more complicated: they may have many dendrites, or observers, looking to see what other neurons are doing. Purkinje cells, which help keep our movements controlled and balanced, can have up to one million dendritic inputs. The silicon neurons of modern AI systems mimic the biology of our brain. They have inputs, they decide when to fire, and they also have neighbours who will react when they see something happening. But comparing a perceptron to a real neuron, beyond abstract similarities, is a basic mistake. The biological node has many more connections, and much more complicated activation functions, than its simulacrum.
And therein lies all the difference. Training LLMs (if training is the right word) somewhat resembles what happens in the real world. Figuring out the rules—what’s connected to what, how many friends need to say they see a light before you turn yours on—is an incredibly expensive and time-consuming process because there are trillions of parameters that need to be tuned (who gets to vote, when do they vote, how long is their vote valid) over many iterations. Before they’re released as a product, they have to predict the next word across billions of sentences. If it guesses incorrectly, a mathematical algorithm calculates the error (called a penalty) and adjusts the internal connections to make a better prediction in the next iteration. This is why LLMs sound so natural, so human. They have “learned” how we speak.
To address these basic flaws in the common perception of LLMs and AI, Doctorow writes:
[The] fears (or hopes) of a nascent superintelligence that will spontaneously arise if we just give enough computing power and training data to large language models are absurd on their face. They make sense only if you believe that “being conscious” is a matter of being really good at guessing which word is statistically most likely to come after the previous one.
He goes on to describe a powerful analogy between throwing memory and compute at invalidated text (and the unsafe assumption that facts will just “ring out” after many iterations), and the still-missing science of cognition and understanding. For this, he invokes the controversial figure of Leland Stanford, the eighth governor of California and later its senator, who founded Stanford University (where I had the privilege of being a consulting professor for a decade beginning in 2007) in honour of his deceased son. Stanford was, according to Doctorow, “a renowned horse breeder, whose ruthless program of culling and breeding produced some of the fastest horses we’d ever seen” in the mid-19th century.
But no amount of equine eugenics was ever going to produce a locomotive. Faster horses are still horses; a different kind of speed required a different kind of machine. That’s the second point, and it’s the one the industry would rather you not make: predicting the next word, no matter how many trillion parameters you throw at it, is still pattern-matching. It is not a different kind of thing happening—not experience, not understanding—just a faster horse. LLMs are incredibly useful, even astonishing. But they aren’t “thinking” about anything. They’re a calculator, with all the heart, soul, perception, and sentience of a hammer.
Finally, we come to the threshold of societal implications, specifically about worker exploitation, and the twin threats of over-investment and mass unemployment before, during, and after the bubble bursts. Here, again, Doctorow draws upon his days as a technology journalist and commercially successful writer of science fiction. His words are vivid, imaginative, instructive, and clear.
Although I take some exception to the general assumption that “all bosses are evil,” there’s no refuting Doctorow’s primary observation, which is that the returns on AI investment—which Goldman Sachs anticipates will be about US$765 billion this year, US$1 trillion in 2027, and US$1.2 trillion in 2028—far exceed the reach of any business model that has been publicly shared. These are unheard-of numbers, in size and speed, and the people writing the cheques expect handsome returns for their risk.
Doctorow explains, with rich examples and accessible detail, that AI—at least in its current form and with our still-limited scientific understanding of what consciousness really is—cannot even achieve human-level performance on some simple tasks. The widely respected cognitive scientist, AI entrepreneur, and bestselling author Gary Marcus concurs. He points out that generative AI systems cannot follow simple instructions, or obey the rules of chess. These are not indictments; they are ordinary limitations of an otherwise useful tool. We don’t reach for a hammer when we need pliers; we don’t expect pliers to decode GPS signals.
But that’s not what the merchants of cognition want employers to believe, and they have trillions of reasons to get us to see it their way. Unfortunately, as Doctorow explains, “Keeping growth alive isn’t about one company or one sector. The entire U.S. economy hangs in the balance.” The circular economy of data centres, chips, and software has created an investment gravity the likes of which we have never seen before. As of June 2026, the profits are just not there, and there is no anticipation of them coming—at least in reciprocal size and speed—that would justify the trillion-dollar investments in the sector.
LLMs can’t replace people, and AI can’t replicate actual thought. But if employers reject these truths and succumb to empty promises and motivated thinking, employees will suffer the immediate, material consequences of being replaced by a tool that cannot actually do their job because their bosses are confused by the perception and temptation of the uncaptured profit margin.
Our Blue Button initiative teaches a lesson that the AI boosters and the AI doomers both keep missing: the question is never what a technology can do in the abstract. It’s who it’s built for, and who pays when the bill comes. We kept a low profile until we were sure, and we didn’t have to stretch any claims because we knew it would work. There was no mystery, and we could explain everything about it. The market liked what it saw. In retrospect, blasting it out there might have limited its adoption. The only people who cared were the ones who benefited from it, and the economic costs were rounding errors in the budget.
As mainstream thinkers join Doctorow and come to recognise the danger, the contours of the profit trap are becoming harder to ignore. We would be wise to heed his warning.
