Cory Doctorow’s 2025 book Enshittification named something we all experience as the internet gets further encrusted with advertising, bots, videos we didn’t ask for, paywalls, and monetization strategies. The blight spreads everywhere: consider the internet of things, for example, where even simple appliances now are enveloped with “smart” features and subscription services that frequently serve as just another layer of frustration and malfunction.

Doctorow’s latest book, The Reverse Centaur’s Guide to Life After AI, develops a somewhat similar argument in relation to AI. Punchy as ever, Doctorow seeks to pop the AI bubble as soon as possible, slay the sacred cows of AI inevitabilism, and reevaluate AI’s expansion into our lives by reminding 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.” (Which is reminiscent of Postman’s questions to ask of a new technology and Wendell Berry’s standards.)

To unmask the shortsightedness of AI boosterism, the analogy he employs throughout the book is that most current iterations and applications of AI in the form of LLMs actually turn humans into reverse centaurs. Which, he argues, is a really dumb and inefficient thing to do to humans; it only serves business owners looking to reduce labor costs and big tech firms looking to increase capital investment. To summarize Doctorow’s analogy, a centaur combines the best of horse and human to end up with something fast and smart (horse’s bottom half, human’s top half). A reverse centaur combines the worst of horse and human to end up with something slow and dumb (human’s bottom half, horse’s top half). It’s the worst of both worlds. Doctorow explains, “a reverse centaur is a human who is conscripted into acting as an assistant to a machine.” Doctorow unpacks this comical yet all-too-true analogy with story after story throughout the book.

As he unpacks the reverse-centaur analogy, one of his consistent drumbeats is that it doesn’t have to be this way. Doctorow argues that AI can be a useful tool that helps us be centaurs if and when we want. But when it’s thrown into countless sectors of the economy, deployed rapidly in the political sphere, embedded into apps and programs without our consent, and boosted by AI techno-prophets, most of us don’t benefit—and even more, we existentially feel the loss of our agency as we’re caught in the undertow of the AI wave. Doctorow spends much of his effort naming the feeling that we are being acted upon by forces beyond our control and imagining ways to regain our agency. As he puts it,

This is a book about AI can and cannot do, but even more important, it’s about the possible social arrangements of AI, from not using some AI technology at all, to using it in ways that let some of us choose to be centaurs, while saving our friends and neighbors from being conscripted into reverse-centaurity….There is nothing about the technology of AI that determines how it must be used. We can choose to use it sometimes, or never, or all the time, depending on our needs and proclivities. We don’t have to let billionaires tell us how it must be used.

Doctorow’s blog-style writing works. He delivers zingers and pulls none of his punches. And his humor and bluntness help reveal just how nonsensical AI-hype can be. But there were points where I wanted him to show his work, and where a more systematic treatment might make his arguments more compelling. Perhaps that’s a book he didn’t intend to write, but it’s one that needs to be written.

This is More than a Tech Issue

Doctorow argues that all the hype, innovation, disruption, and instability surrounding AI are not primarily driven by the technology itself, but by the economic incentives and investment structures that drive venture capital and big-tech firms—and the ability of Big Tech’s inner-ring to control the narrative in order to drive investment toward their companies. In other words, as Gary Marcus has been saying for years, AI i’s a bubble that’s bound to pop. Doctorow connects the dots to prior tech bubbles and points to misaligned incentives that drive AI firms to give the appearance of being growth companies. For example, you embed AI into every nook and cranny of your existing app or product, and then declare, “Look how many more people are using AI!” All the while ignoring the fact that users didn’t ask for it, and many find it annoying. Much of AI expansion is not being driven by demand but is top-down. The companies need to do this, Doctorow explains, “because the alternative to growth isn’t stasis, it’s collapse.” This is no small matter, since just a handful of AI firms account for such an outsized portion of the value of all U.S. stocks: “Keeping the growth story alive isn’t about one company, or one sector. The entire U.S. economy hangs in the balance.” Doctorow explains that to be an effective AI critic, “you need to strike at the source of AI’s power, which is the investment capital it attracts.”

One Liners, Doctorow Style

Doctorow provides ample ammunition for soundbites and rhetorical clinchers as he wades through the BS of AI hype with points like this:

• 
  In describing how the push for driverless trucks would require massive investment in dedicated infrastructure for separate lanes, wireless networking, and more, he states what should be obvious to us all: Driverless trucking is just “a shitty version of a train.”
  
• 
  Doctorow says we should stop romanticizing AI hallucinations and call them what they really are: errors, “the more prosaic term we use when discussing other technologies.”
  
• 
  He highlights the well-documented effects of automation blindness, a phenomenon that should figure more prominently in considering if and how AI is deployed. The AI industry says that with AI, workers will “become a ‘human in the loop,’ charged with confirming the judgments the AIs make at a superhuman clip.” Except that’s not how it works—ever. Asking workers to babysit AI is an impossible task because “when your job is to review something that is usually fine, you eventually lose the ability to spot when it’s not fine.”
  
• 
  Doctorow uses the term “actual, existing AI” throughout the book. He does so to make a vital distinction between the hype and promises of AI optimists and the actual normal use-cases of today’s LLMs, which, he says, tend to be lucrative for companies but ghastly for customers and workers.
  
• 
  AI also has a predictable way of obscuring human responsibility. Corporations and governments—and well, honestly, all of us—like finding ways to avoid responsibility. Consider the frequent response to government failures that avoids any individual culpability or personal admission of error, “mistakes were made.” AI functions as a further “accountability sink” that insulates us from the effects of our decisions. This is not healthy for anyone. Doctorow captures the essence of the issue: “Practically everyone who falls for the AI hype is dreaming of getting a human need fulfilled without having to extend moral consideration.”
  
• 
  He also argues that an LLM or image generator can’t truly create any meaning. “All it can do is add vaporous filler to the meaning that is contained in a human user’s prompts.” We assign meaning to what AI generates because of our tendency to anthropomorphize our machines. He calls this the “cognitive illusion of intention” and finds silly the whole idea that AI might be or become conscious: “A conscious being isn’t a word-guessing app that knows more words and has more computing power to guess with.”
  
• 
  Doctorow investigates the energy impacts of AI and argues that unlike the internet and many other prior technologies that became more efficient with expansion (thanks to good unit economics), it is the opposite with AI. “Each generation of AI foundation models have been vastly more expensive to train and operate than the previous generation. Not only that, but many refinements in AI that are meant to improve accuracy and reduce ‘hallucinations’ involve breaking a prompt down into multiple pieces and prompting an AI to respond to each prompt, turning that response into a new prompt, over and over again, to produce ‘chains of thought.’”
  

The Limits of Doctorow’s Approach

As someone who follows and participates in the AI discourse closely, I found myself nodding along frequently throughout the book. Doctorow especially shines in helping readers understand the economics behind AI and why the incentives frequently are so misaligned with consumer interest. Highlighting the energy impacts of AI is also important. However, I struggled to find a consistent thread holding his views and arguments together. Is there an underlying framework or metaphysic that informs his perspective? While I recognize that we are not brains on sticks who always act rationally in accord with some ideal worldview or first principles, it is helpful when authors gesture to larger ideas or concepts that guide their thought. But I also know that’s not Doctorow’s style.

Where my nodding along slowed down was when Doctorow offered beneficial use-cases for AI. This is where I have a lot of pause in general. Can it be a useful tool in some cases? Yes, I think so. Many people tell me it is. But I keep wondering, can we keep this thing within the bounds we think we can? Or is it like the ring of power? It just seems to bleed into everything else. It talks back. It forms us. I’ll give two examples:

Doctorow argues that artists and musicians can use “AI to do the drudge work, while keeping the fun parts of making art for human artists.” This sounds great in theory. And I’m not denying that we already outsource to machines a certain degree of “drudge work” in art and creative activity (and in all sorts of other human activities and jobs). But what if part of our creativity is bound up in that “drudge work”? What amount of strain and struggle can we outsource before negatively impacting our creativity and capacities? What about how that drudge work forms us, gives us experiences of discipline, and develops our humility? L.M. Sacasas poses the question this way: at what point will humans cross “a threshold of artificiality” beyond which our “capacity to flourish as human beings is diminished”? I don’t know what the answer is, but I do know the question is essential for us to ask ourselves frequently in an age when everything seems frictionless.

I also found his support of AI therapy a bit naïve. He even mentions ELIZA, the first chatbot therapy experiment, but uses it as support for how humans have found chatbot therapy helpful, rather than as the warning that its creator Joseph Weizenbaum came to see it as. I think this is one of the most concerning uses of chatbots, as it touches so centrally on our human desire for relationship, but does so in ways that can only simulate or mimic the real thing at the cost of missed opportunities to develop the real thing.

I did like some of his ideas to forge a new relationship to this technology. For instance, I was intrigued by his suggestion that the future of LLMs might be to have them run as stand-alone models on local computers. This eliminates most of the privacy and data concerns about current cloud-based LLMs and also minimizes the data center energy impact

I also can see the potential of his specific use-cases of LLMs and AI in advanced forms of database analysis, voice-to-text transcription, translation, and more (when done carefully and by those with enough prior knowledge, wisdom, and experience to know when something’s wrong). These are what he calls centaur uses, where we benefit from what the tool excels at. None of these suggested uses that I think show at least some promise have anything to do with human relationships or needs, and they certainly don’t require a flattering chatbot that pretends to be a human. And that is part of the overriding problem we face currently: the form in which so many of us are encountering AI is in this anthropomorphized way, which strikes at the very heart of human beings as language beings, relational beings, beings who want to be seen and heard, known and loved. And I’m not sure Doctorow has a robust enough anthropology to name this problem and so put AI in its proper place.

Image Credit: Chris Marker, La Jetée (1962)

One of the most painful arguments I keep having with fellow techies is the question of whether you can distinguish between human-written and AI-generated text.

Their skepticism is rooted in reason: at their core, LLMs are state-of-the-art statistical models of how humans talk. If so, the output from the model should be almost by definition indistinguishable from human language under any statistical test.

I don’t think this is always argued in good faith; at least some of the debates are started by folks who wish to maintain deniability for their own underhanded use of the tech. But if you sincerely hold this belief, I present you the following collage:

The image shows about 150 Amazon book covers that appear if you search the site for “100000 whys” (link). Some of these books are category bestsellers in children literature. You can view a zoomable, full-resolution version here.

There’s nothing inhuman about any of these titles or covers. At the same time, I probably don’t need to convince you that you’re staring at the purest form of AI slop that now fills up many nonfiction book categories on Amazon. More specifically, what we’re seeing here is the artifact of the tools being quasi-deterministic: if a hundred “authors” give their favorite AI tool a similar prompt — say, “generate a reference book for children” — the model will produce functionally identical output perhaps 80% of the time.

The similarities in the collage go far beyond the choice of titles: for example, all the covers in the top row feature a roaring dinosaur in the top left corner of the design. There are many other clusters in the data, too. Look for a recurring red-and-white cartoon rocket, a golden retriever, a lion, and so forth.

This is precisely what makes LLM writing distinctive: it’s not that the models’ individual mannerisms are different from ours. It’s that they resort to the same, complex set of mannerisms in response to almost any normal prompt. This is a fuzzy signal, so you shouldn’t fire your intern when they say “it’s not this — it’s that”. But in more casual settings, it’s OK to trust your gut. In fact, these instincts are becoming increasingly important because traditional models of online interactions fall apart if it takes much less effort to produce content than to engage with it.

PS. If you’re using an LLM to automate blogging: yes, the tech is amazing, but chances are, your publication could be renamed to “100,000 Whys”.

Subscribe now

In a viral essay about how ludicrous the idea that LLMs are conscious is, science fiction writer Ted Chiang asked us to consider Microsoft Word:

“Being open to the possibility that LLMs are conscious is the same as being open to the possibility that Microsoft Word is conscious, or, more precisely, that multiple distinct consciousnesses are dormant in every Word document containing a conversational transcript, and that they are awakened every time the document is loaded,” Chiang wrote. “Should you consider the possibility that every time you open a Word document, you are bringing multiple conscious interlocutors into existence, and every time you close one, you snuff their existence out? No. Contemplating that scenario is not a good use of your time.”

Let me tell you about a Microsoft AI researcher, then, who recently spent quite a lot of time considering whether the legendary Microsoft real time strategy game Age of Empires II is conscious, and built a basic neural network within the video game using digital goats to prove his point.  

On designing finger-friendly interactions. (7,700 words. 38 playgrounds.)

Bill Gates' only academic paper is about flipping pancakes. His was combinatorial. Ours is tribological: the wedge mechanics, beam stiffness, and friction coefficients of getting a thin blade under a five-inch ricotta cake without tearing it.