If you'd like an essay-formatted version of this post to read or share, here's a link to it on pluralistic.net, my surveillance-free, ad-free, tracker-free blog:
https://pluralistic.net/2026/08/03/andor/#either
The greatest magic trick of them all is lying. The reason you can't figure out that coin vanish even after the conjurer performs it three times in a row is that they didn't do the same trick three times in a row! They did three different tricks: "Didn't catch it? Here, let me do it again!" is a lie:
There's times when it makes sense to treat two outcomes as the same, even if they were produced by very different means. As a reader, my enjoyment of your novel is the same whether it was dictated, typed on an Underwood Noiseless, keyed into a word processor, or scratched out with a fountain pen:
There's plenty of routes that arrive at the same place, and if the destination is all that matters to you, it's fine to ignore the journey. But often, those end-points have subtle differences that are only revealed when things go wrong. If all you care about is how things work, chances are good that you're in for an unpleasant surprise when things fail.
I recently found myself arguing with an interviewer about whether AI is, or could be, conscious. We weren't arguing about whether it might someday be possible to make an artificial consciousness â as a materialist, I'll happily stipulate to this. I think that everything we call "consciousness" is the result of a physical process occurring within our bodies (and possibly around them?), so I think it's perfectly reasonable to imagine that someday we might create another physical process that produces the same effect.
But that's not what the interviewer wanted to argue about. His point was that teaching more words to the word-guessing program would produce consciousness, an argument I always liken to "breeding horses to run faster and faster until one of them foals a locomotive." In support of this (outlandish) proposition, the interviewer performed a kind of cognitive coin-trick: "I can often predict what my wife is going to say, and so can a chatbot that's been trained on her words. Therefore, we're both doing the same conscious work â and therefore the chatbot will eventually be as conscious as I am."
"Predicting what you will say through an understanding based on a theory of your mind" and "predicting what you are going to say based on a statistical analysis of your utterances" might produce the same outputs, but they are not the same trick. You can tell by what happens when the trick fails.
My wife and I have been together for 23 years now, and there's plenty of times that we can finish each other's sentences â and so can the autocomplete on our phones. The autocomplete manages the trick by exploiting the fact that we often repeat ourselves. But we manage the trick by understanding each other (and by exploiting the fact of repetition).
When my wife says something surprising â because she is angry or delighted, sad or happy â I can make a reliable guess about what caused my prediction to misfire. Our "sentence completion" trick doesn't emerge from a rough, automatically generated mental table of the statistical likelihood that word A will follow word B. We also understand why those combinations appear in each other's speech and writing.
"Understanding" and "statistical extrapolation" can often lead to the same place, but when they don't, "understanding" provides a way forward, while "extrapolation" founders. Both work fine, but only one fails gracefully. The two tricks only appear the same, but they are fundamentally different.
AI's investor story â and the science fiction tales of AI's eventual capabilities that underpin that investor story â makes heavy use of this conjurer's trick, in which two different outcomes are equated to one another because they resemble each other.
This "ignore the journey, focus on the destination" idea is baked very deeply into the way we think about AI. Take the "Turing Test," a complicated and nuanced thought-experiment proposed in 1950. Over the ensuing 75 years, Turing's thought-experiment has been stripped down into a blunt metric: "Can a chatbot trick a human into thinking it is also human?"
"I mistook a chatbot for a human" and "I took a human for a human" arrive at near-identical places, but they are subtly and importantly different. The erroneous assumption that my phone's autocomplete is actually a person who understands me well enough to finish my sentences works fine, but the instant I turn to it for understanding, it will fail very badly. Autocomplete's predictions are always grounded in who you used to be, which means autocomplete knows very little about who you are now, and absolutely nothing about who you will become:
https://reallifemag.com/instant-recall/
The low-rez Turing Test that captured popular discourse is profoundly misleading. It's the unsound foundation of a worldview that renders you incapable of distinguishing your understanding of your spouse from their phone's autocomplete function. It's the self-serving rationale that leads you to declare yourself a proud stochastic parrot:
The AI bubble is (seemingly) full of contradictions, but â like those baffling coin-tricks â these contradictions often resolve themselves very neatly once you realize that the "contradiction" is actually just two things that appear to be one.
For example, some of the billionaires who put up the first several hundred million for AI are solipsists who just don't believe other people are entirely real and therefore find it easy to believe that AI can do their jobs. Other billionaires are cynics who think that bosses can be sold defective worker-replacing chatbots because they're credulous suckers for that pitch, the same way they believe that desperate young men are suckers for Joe Rogan's useless and/or dangerous supplements and peptides:
Billionaire AI true believers and billionaire AI cynics make for a powerful coalition. The roadblocks that might discourage the first group are easily hurdled by the second, and vice-versa. You don't have to believe AI works to believe it can be sold, and you don't have to be motivated by the sales opportunity to believe that AI is about to become god.
Almost every debate I get into about AI turns out to be an unjustified, unacknowledged conflation of two things that seem similar, but have profoundly different underlying characteristics. Take this argument: "Every time we extend rights to the nonhuman world â watersheds, endangered animals, ecosystems â the world gets better. Let's extend rights to AI â whether or not we think it's a 'person' and so reap those benefits."
This, too, is a coin trick. Extending rights to nature reliably makes the world better, but extending rights to constructs makes the world far worse (Exhibit A is corporate personhood) (obviously).
A few moments' thought reveals the difference. If we extend rights to a watershed, that might result in an AI data-center being killed. If we extend rights to AI, that might lead to sacrificing the watershed to cool the data-center:
Then there's AI and labor. The world is full of skilled workers who have found ways to use AI on the job that they insist have improved their work. It's also full of skilled workers who warn us that on-the-job AI is producing tech debt at unimaginable scale, seriously depreciating the quality of the tools we use today, and setting us up for painful reckonings in the future.
This (seeming) contradiction melts away once you realize that these workers only appear to be doing the same thing. The first group of workers, excited about their AI-assisted output, are "centaurs": people assisted by machines; workers who choose the time and manner of their AI adoption. The second group are "reverse centaurs": people recruited to serve as peripherals for machines, who direct their actions and workflow:
Note that this isn't the same thing as saying "A skilled worker who adopts a tool willingly is always right and will produce a better output as a result." Nor is it saying, "The tool is so flawed that workers who claim it works for them must be deluded."
That's another coin trick! The reality â again â is that this is two things: some workers whose AI-assisted work is measurably worse are wrong about AI making their work better (centaurs, but wrong), and; some workers are being forced to use AI and know damned well that it's making their work worse (reverse centaurs).
Finally, there's an economic coin-trick: "AI will destroy jobs." Sure, yes, AI is destroying jobs. But there's a vast difference between "You got fired because an AI can do your job" and "You got fired because your boss was convinced that the AI can do your job, even though it cannot."
This is one of the most consequential coin-tricks, because it's a real convincer for the investors who are funding the AI bubble. The difference is huge: "AI can do your job" means you're well and truly screwed. If an AI can really replace a contract lawyer, then everyone who needs a contract written or evaluated should be on the side of mass technological unemployment for contract lawyers. The point of contract lawyers is to produce contracts, not to pay contract lawyers' law-school debts and mortgages.
BUT! If some BigLaw's credulous partners can be suckered into firing their juniors and replacing them with chatbots who bill you $1,200/hour to produce unenforceable, error-riddled contracts, then everyone who needs a contract is on the same side as the contract lawyers â united in opposition to their bosses:
Every time we fail to draw this distinction, we help an AI boss raise another billion dollars. Every time we insist on this distinction, we hasten the day that the AI bubble pops, thus sparing a few more everyday savers and innocent bystanders from being wiped out in the crash we can all see on the horizon:
As "Cathy" so aptly put it: "The thing that is a good tool for the skilled people is being sold as a thing to reduce the number of skilled people hired":
The former is a normal technology. The latter is the root of a catastrophic folly that is destroying our environment, destroying workers' lives, destroying the quality of the goods and services we rely on, and which will shortly destroy our economy.
Anya is live and ready to show you everything. Watch her strip, dance, and perform exclusive shows just for you. Interact in real-time and make your fantasies come true.
â Live Streamingâ Interactive Chatâ Private Showsâ HD Qualityâ Free Actions
Free to watch ⢠No registration required ⢠HD streaming
ITHACA and NYC! I'm heading your way for a zillion events from Sept 11-17. Here's a list of open-to-all CORNELL activities including two major keynotes; a movie night with dinner and discussion; and a public event at CORNELL TECH in NYC. I'm also appearing at BUFFALO STREET BOOKS on Sept 11 and at AUTUMN LEAVES BOOKS on Sept 13. I'll also be at the BROOKLYN BOOK FESTIVAL on Sept 21:
The most ENSHITTIFICATION-PROOF way to get the Enshittification audiobook, ebook and hardcover is to pre-order them on my Kickstarter! Help me do AN END RUN around the AMAZON/AUDIBLE AUDIOBOOK MONOPOLY and DISENSHITTIFY your audiobook experience in the process.
My latest Locus column is "Reverse Centaurs," and it sets out to unravel a paradox: how is it that some AI's users describe their experience as a hellish ordeal, while others delight in the ways that AI is changing their lives for the better?
A "centaur" is a human being who is assisted by a machine (a human head on a strong and tireless body). A reverse centaur is a machine that uses a human being as its assistant (a frail and vulnerable person being puppeteered by an uncaring, relentless machine).
Let me give you an example: remember at the start of the summer, when Hearst published a summer reading guide that was full of nonexistent books that had been "hallucinated" by a chatbot?
But in a followup story, Koebler noticed something that the first round of dunks and memes about this poor guy had missed: this same writer had his name on many of these "best of the summer" lists in this supplement. He was practically the sole author of an entire 64-page insert:
And that's where it gets interesting. Koebler got his start in journalism as an intern at the Washington Monthly, where he worked on lists like these:
https://www.404media.co/podcast-ai-slop-summer/
When Koebler was doing this work, he'd be part of a team of three interns, overseen by an experienced journalist, backstopped by an extensive fact-checking department. Those little lists take a surprising amount of work, if you really care about their quality.
The freelance writer who authored this giant summer reading guide with all its lists had been tasked with doing the work of literally dozens of writers, editors and fact-checkers. We don't know whether his boss told him he had to use AI, but there's no way one writer could do all that work without AI.
In other words, that writer's job wasn't to write the article. His job was to be the "human in the loop" for an AI that wrote the articles, but on a schedule and with a workload that precluded his being able to do a good job. It's more true to say that his job was to be the AI's "accountability sink" (in the memorable phrasing of Dan Davies): he was being paid to take the blame for the AI's mistakes.
He was, in other words, a reverse centaur.
Now, I am a freelance writer as well, and not so long ago, I wanted to quote something smart I'd heard on a podcast in an article, but I couldn't remember where I heard it. So I downloaded Whisper, an open source AI transcription model from Openai, to my laptop. I threw the last 30 hours' worth of audio that I'd listened to at it, and worked away on other stuff for an hour or two. When I checked again, I had a folder full of pretty reliable transcripts. I searched the text, found the quote, and opened the audio to the supplied timecode to double-check it. I was a centaur. I got to decide how to use the AI, and I only had to use it in ways that made my work better and more satisfying.
This, I think, is the explanation for the paradox of AI: the AI users who are being immiserated and precaratized by bosses who have been convinced to fire their colleagues and pile their work on the terrorized survivors of the layoffs hate the AI, because it makes their life worse in every way.
Whereas the people who choose when and how to use AI â the centaurs â are only using AI to the extent that it is useful, and throwing it away when it's not. They may make poor choices about the AI, but those choices are theirs, they are not imposed from on high. A bicyclist who chooses to commute on two wheels can have a glorious ride, or they can ride like a maniac and end up eating dirt, but they are having a fundamentally different experience from, say, a gig delivery platform rider who has been given an impossible quota and is having their pay eroded by algorithmic wage discrimination:
I was very happy to put this analysis in the pages of Locus, the trade magazine for the science fiction field. The job of a science fiction writer is only incidentally to describe what a technology does â at its best, science fiction interrogates who the technology does it to and who the technology does it for.
This is a political act of resistance. Margaret Thatcher's motto, after all, was "There is no alternative," by which she meant, "Stop trying to think of alternatives." The bully's trick is to present your defeat as a fait accompli: "Resistance is futile."
Tech bosses practice a form of vulgar Thatcherism all the time: Mark Zuckerberg wants you to think there's no way to talk with your friends without letting him listen in; Sundar Pichai wants you to think there's no way to search the web without being spied on; Tim Cook wants you to think there's no way to have a safe and reliable computing experience without giving him a veto over which software you install; Satya Nadella wants you to think there's no way for you to edit a Word file without letting your boss compare your keystrokes-per-minute to your co-workers:
And AI bosses want you to think that the only way to use these tools is to displace and immiserate labor, because that's the promise they raise investment capital on:
AI is a bubble. If it wasn't a bubble â if it was just a bunch of computer scientists and product teams tinkering with possible uses for advancements in back-propagation, generative adversarial networks and machine learning â there wouldn't be any controversy here. A programmer who uses a chatbot to autogen a bunch of cross-browser CSS stylesheets that mostly work, after some tinkering, would maybe mention that fact over beers â but they wouldn't get sucked into a cult obsessed with outlandish scenarios in which the chatbot wakes up and turns us all into paperclips:
AI is a bubble. Bubbles burst. We're in for a near-total collapse of the AI investment mania. Most of these companies will fail. Many planned data-centers will never be opened. Many existing data-centers will be shuttered. When that happens, what will be left?
AI is a bubble, and when bubbles burst, they sometimes leave behind a productive residue. At home, I enjoy 2GB symmetrical fiber optic internet, because AT&T was able to light up some of the dark fiber that Worldcom fraudulently raised billions for. Worldcom's CEO died in prison after scamming the finances of ordinary people, and the world would be a better place if that had never happened, but there was some productive residue left behind, and many of us are reaping the benefit today:
Contrast that with the cryptocurrency bubble. When that bursts, we'll still have a smattering of programmers who've had a subsidized education in cryptography and secure programming in Rust, but mostly what crypto will leave behind is bad Austrian economics and worse monkey JPEGs. Like Enron, crypto will leave nothing much behind of any value.
All bubbles are bad, but some are more productive than others. When the AI bubble bursts, there will be stellar bargains on GPUs (it would be ironic if scientists snapped them up at pennies on the dollar and used them for climate modeling). We'll have a lot of technical people who are much better at applied statistics than they were a decade ago. And there will be the open source models, like Whisper, the tool I used to transcribe all those podcasts.
These open source models run on commodity hardware, and while the climate costs of creating those models is terrible, they're here now, and operating them isn't especially energy-intensive. When I used Whisper to transcribe 30 hours' worth of podcasts, my laptop's fan didn't even switch on.
What's more, open source hackers are doing amazing things with these tools â far more than the giant corporations that released them ever anticipated. These "toy" models were released as a way to entice programmers into specializing in cloud systems operated by the big tech companies, but it turns out that these standalone models can do amazing things, and aren't just a demo for a big, doomed foundation model:
It doesn't matter what happens to Openai; Whisper is here to stay. It's already being rolled into other standard tools â the latest version of ffmpeg integrates Whisper and can autogen captions:
The things these open source standalone models can do will only expand, and they will become a given for our computing applications. Your computer or phone will be able to transcribe audio and do cool image-editing stuff like erasing strangers from the background of a photo as a standard feature.
That's the good news. The bad news is all the damage the bubble is doing now and all the further damage that will come from its collapse. Today, we're getting the climate impact, obviously, and the immiseration of all those workers who are being reverse-centaured by an AI that can't do their job, but whose manufacturer's salesforce convinced their boss to fire them and replace them with an AI anyway.
After the bubble bursts, there will be the mass incineration of everyday people's retirement savings and the knock-on effects as the whole market craters. And long after that, there will be the terrible impact on our society's ability to do things, as defunct foundation models grind to a halt, after the people they replaced are long gone and can't step in to pick up the work they fumble. We are busily filling the walls of society with digital asbestos and we'll be digging it out for generations to come.
Every day the bubble persists, the harms of today and tomorrow increase. We need to burst that bubble as soon as possible. That's how I came to spend the summer writing a book for Farrar, Straus and Giroux with the working title The Reverse-Centaur's Guide to AI, whose goal is to improve the quality of AI criticism so that it inflicts maximum damage on AI swindlers and their terrible investment bubble.
It'll be out in 2026, but for now, you can have a look at my Locus column:
Click here to pre-order my next book, ENSHITTIFICATION: WHY EVERYTHING SUDDENLY GOT WORSE AND WHAT TO DO ABOUT IT
If you'd like an essay-formatted version of this post to read or share, here's a link to it on pluralistic.net, my surveillance-free, ad-free, tracker-free blog:
I'm on a 20+ city book tour for my new novel PICKS AND SHOVELS. Catch me in SAN DIEGO at MYSTERIOUS GALAXY on Mar 24, and in CHICAGO with PETER SAGAL on Apr 2. More tour dates here.
AI can't do your job, but an AI salesman (Elon Musk) can convince your boss (the USA) to fire you and replace you (a federal worker) with a chatbot that can't do your job:
If you pay attention to the hype, you'd think that all the action on "AI" (an incoherent grab-bag of only marginally related technologies) was in generating text and images. Man, is that ever wrong. The AI hype machine could put every commercial illustrator alive on the breadline and the savings wouldn't pay the kombucha budget for the million-dollar-a-year techies who oversaw Dall-E's training run. The commercial market for automated email summaries is likewise infinitesimal.
The fact that CEOs overestimate the size of this market is easy to understand, since "CEO" is the most laptop job of all laptop jobs. Having a chatbot summarize the boss's email is the 2025 equivalent of the 2000s gag about the boss whose secretary printed out the boss's email and put it in his in-tray so he could go over it with a red pen and then dictate his reply.
The smart AI money is long on "decision support," whereby a statistical inference engine suggests to a human being what decision they should make. There's bots that are supposed to diagnose tumors, bots that are supposed to make neutral bail and parole decisions, bots that are supposed to evaluate student essays, resumes and loan applications.
The narrative around these bots is that they are there to help humans. In this story, the hospital buys a radiology bot that offers a second opinion to the human radiologist. If they disagree, the human radiologist takes another look. In this tale, AI is a way for hospitals to make fewer mistakes by spending more money. An AI assisted radiologist is less productive (because they re-run some x-rays to resolve disagreements with the bot) but more accurate.
In automation theory jargon, this radiologist is a "centaur" â a human head grafted onto the tireless, ever-vigilant body of a robot
Of course, no one who invests in an AI company expects this to happen. Instead, they want reverse-centaurs: a human who acts as an assistant to a robot. The real pitch to hospital is, "Fire all but one of your radiologists and then put that poor bastard to work reviewing the judgments our robot makes at machine scale."
No one seriously thinks that the reverse-centaur radiologist will be able to maintain perfect vigilance over long shifts of supervising automated process that rarely go wrong, but when they do, the error must be caught:
This is bad enough when we're talking about radiology, but it's even worse in government contexts, where the bots are deciding who gets Medicare, who gets food stamps, who gets VA benefits, who gets a visa, who gets indicted, who gets bail, and who gets parole.
That's because statistical inference is intrinsically conservative: an AI predicts the future by looking at its data about the past, and when that prediction is also an automated decision, fed to a Chaplinesque reverse-centaur trying to keep pace with a torrent of machine judgments, the prediction becomes a directive, and thus a self-fulfilling prophecy:
AIs want the future to be like the past, and AIs make the future like the past. If the training data is full of human bias, then the predictions will also be full of human bias, and then the outcomes will be full of human bias, and when those outcomes are copraphagically fed back into the training data, you get new, highly concentrated human/machine bias:
By firing skilled human workers and replacing them with spicy autocomplete, Musk is assuming his final form as both the kind of boss who can be conned into replacing you with a defective chatbot and as the fast-talking sales rep who cons your boss. Musk is transforming key government functions into high-speed error-generating machines whose human minders are only the payroll to take the fall for the coming tsunami of robot fuckups.
This is the equivalent to filling the American government's walls with asbestos, turning agencies into hazmat zones that we can't touch without causing thousands to sicken and die:
If you'd like an essay-formatted version of this post to read or share, here's a link to it on pluralistic.net, my surveillance-free, ad-free, tracker-free blog:
This is truly the opposite of steampunk. Somehow, our bosses have invented a form of craft-labor â where you work from your own vehicle or home, using equipment you pay for â that has all the humiliations, dangers and petty authoritarianism of the industrial factory.
This is the worst of both worlds. Under the New Deal, factory workers teamed up with progressive regulators to force the owners of giant factories to share the efficiency gains of the assemblyline, creating the âLarge-Firm Wage Premiumâ (where workers at big companies made more money, not less).
Today, the large-firm wage premium is dead. Workers are moving out of the factory, back into their homes (and cars), but those homes and cars are being transformed into factories, thanks to the camera- and mic-studded digital devices that monitor workers more closely than even the meanest, pettiest foreman could.
It neednât be this way. The Luddites presaged the steampunks, imagining technology to liberate, not to enslave. Technological tools could be labor organizersâ secret weapons, shifting power back towards workers.
Anya is live and ready to show you everything. Watch her strip, dance, and perform exclusive shows just for you. Interact in real-time and make your fantasies come true.
â Live Streamingâ Interactive Chatâ Private Showsâ HD Qualityâ Free Actions
Free to watch ⢠No registration required ⢠HD streaming
This ethic of technophilia, labor autonomy, solidarity and loose coordination was beautifully summed up in the motto for Magpie Killjoyâs wonderful Steampunk Magazine:
If sf asks, âwhat if the machine had a different social arrangement?â then steampunk asks, âwhat would it be like if we could have the productive benefits of machines without their regimentation?â
A craft worker enjoys enormous autonomy. If they get a cramp or need a bathroom break, they can just stop. If theyâre hungry, they can eat. If the landscape outside the window is looking especially picturesque, they can stop and contemplate it, or even step out into the fresh air to enjoy it.
Even the most labor-friendly, cooperatively owned assembly line canât function if its workers do their own thing. The price of factory efficiency is autonomy: a worker in a multi-stage process has other workers upstream and downstream depending on them to maintain the pace and regimentation of the line.
Steampunk is fantasy in that it imagines lone craftspeople working with all the autonomy of the individual mad scientist inventor or tinkerer, but producing goods characteristic of the factories where workers had to check their autonomy at the door.
Thatâs a utopian vision, one that was especially enticing in the 2000s, when internet collaboration tools allowed thousands of strangers to engage in large, collective endeavors, like writing an encylopedia or an operating system, without any bosses, working at their own pace, relying on version control systems and wiki pages to coordinate their labor while they worked their tools in their craftersâ cottages all over the world.
The factory owners who built their âdark, Satanic millsâ werenât interested in making life easier for textile workers by automating their labor. They wanted to make workersâ lives harder.
Textile machines were valued because they were easier to operate than the hand-looms that preceded them, and that meant that workers who wanted a fair wage for a fair dayâs work could be fired and replaced with new workers, without the logistical hassle of the multi-year apprenticeship demanded by the hand-loom and its brethren.
As Brian Merchant documents in Blood in the Machine, his stunning, forthcoming history of the Luddites, the factory owners of the industrial revolution wanted machines so simple that children could work them, because that would let them pick over Englandâs orphanages, tricking young kids to come work in their factories for ten and twelve hour days.
These children were indentured for a period of ten years, starved and mercilessly beaten when they missed quota. The machines routinely maimed or killed them. One of these children, Robert Blincoe, survived to write a bestselling memoir detailing the horrifying life of the factory ownersâ child slaves, inspiring Dickens to write Oliver Twist.
