Is There Life After AI? Cory Doctorow Thinks So.

The artificial intelligence boom is fueled by an elite wager that software can supplant expensive human expertise. The tech industry and its investors are pumping hundreds of billions of dollars into the technology — Goldman Sachs forecasts that global AI-related investment will soon exceed $1 trillion. At the same time, government regulatory regimes are inadequate and consumer use cases often vague, misleading, or utterly counterproductive.

The economics of AI don’t make sense unless one factors in what its backers hope to gain: a way to displace workers. In his most recent book, The Reverse Centaur’s Guide to Life After AI, Cory Doctorow explains how people have become tools of the machines and the capitalists who own them.

Doctorow recently spoke to Jacobin to discuss the rise of AI, the financial conjuring tricks behind the boom, the copyright fight, what it all means for workers — and why the answer to AI doomerism lies in building workers’ rights and public power.

David Moscrop

People talk about AI all the time. You get through the conversation and realize no one has specified what they’re talking about. So, I want to start with the simplest, most complicated question. What is AI?

Cory Doctorow

That question implies that you understand the most important thing about AI, which is that we call something different “AI” every five to ten years, and we’ve been doing that since the mid-1950s. You could say AI is the label we apply to anything that is doing something that we currently believe is the unique province of a conscious mind. Then it stops being AI once we realize that it’s doing that without being a conscious mind, and we apply the label to something else.

But this new, specific kind of AI is actually really useful and interesting from a class-struggle perspective. It’s useful to understand what this AI does differently from the AIs that preceded it. Before this version of AI, we had something we called symbolic AI. The way it worked was that you’d build a software model of the thing that you wanted to use machine learning on. If you wanted to use machine learning to figure out how to play chess, you would first build a software model of a chess board, chess pieces, and chess rules. Then you would use machine learning — a statistical technique that involves analyzing previous games, running iterations, and asking for human feedback — to try to improve the chess game that the AI could play.

What the current version of AI does, this theory-free version, is it swaps out these models of the world for a brute-force approach. This is why we call it theory-free. We take all the stuff in the world we can observe, and we feed it to all the computers we can find, and we ask the computers to find statistical relationships between things to try to understand whether A causes B. And while we all know correlation is not causation, causes and effects are correlated.

The problem of correlation and causation is that you might think A causes B, but actually it’s A and A’ that cause B. But A and A’ happen at the same time 99.9999 percent of the time, so we can’t distinguish between them. The answer with AI is to keep adding training data until there are enough examples, even with a thing that only happens one in a million times, that we can actually see that it’s A and A’ that’s causing B and not just A. And we just keep throwing more computing power at it.

The weird thing is that we do this for things where we actually know how they work. We know how to play chess, but we ask LLMs to play chess for us. The problem is that, because they’re theory-free, they don’t know what a chess board is. If you ask an LLM to play chess, it will sometimes put a piece on a square where it already has a piece because it doesn’t know that it’s playing chess. It’s just pulling statistical guesses out of a multidimensional array that tries to represent the relationship of every chess move to every other chess move in the training set. So that’s what is going on here.

We’re also using it to try and pull out statistical relationships between things whose causality we don’t understand. There are far more things in the world that we don’t understand than those we do, but we can observe what seems like a causal relationship.

If you get sick tonight and go to the doctor, you might get a medicine whose method of action is not understood. But we know it works, so we give it to people who are sick. We don’t always wait until we understand it before using it — although we do study it.

But here’s the thing: by building a system where we can swap brute force for understanding — so, money for theory — we clear the very last hurdle for finance capital as a source of power. Expertise and money are both unevenly distributed, but money is collected in the hands of people who don’t have much expertise, while expertise is collected in the hands of people who don’t have much money.

There is a durable source of power among experts. If you don’t want to work for Arthur Sackler, he can’t make you. And if you’re the only person who understands some abstruse area of pharmacology, Arthur Sackler can’t do research in that area. But Arthur Sackler is very, very rich. When you swap expertise for money, you don’t need to convince the only pharmacologist who understands some narrow area to come work for you. You can just throw money at the problem.

This is why the AI industry raised a trillion dollars this year on $50 billion revenue: because the finance sector is in love with the idea of a technology where money can be substituted for expertise.

The Brief Life of the Big-Firm Wage Premium

David Moscrop

In your latest book, you describe this dynamic as a reverse centaur — “a machine that uses a human being as its assistant” instead of the other way around. How did we end up on the other half of this thing? Can we flip it back around?

Cory Doctorow

That dynamic is intrinsic to this kind of AI, in the sense that this kind of AI is really, really expensive and only starts to make money if you can displace a lot of high-wage labor. And this is another area where we have to be really careful to be good materialists and distinguish the sizzle from the steak. Pauperizing commercial artists, pornography performers, and taxi drivers is not a way to recoup the expense associated with training AI. These companies are not doing this to get rich. They’re doing it to demonstrate how awesomely powerful the AI is.

No language on earth contains the phrase “as rich as a taxi driver.” Replacing Uber drivers who supply the cars, insurance, depreciation, fuel, and the cost of shampooing vomit out of the backseat with expensive robots that you own doesn’t pencil out — unless your goal is to say, “If we can replace taxi drivers, imagine what we can do to accountants, lawyers, doctors, teachers, radiologists, nurses. And especially people who work in fields where you are legally obliged to tell your boss to f-ck off if they tell you to commit a crime, like medicine, finance, education, and so on.”

That’s what has them so excited. We got here because they saw an opportunity — in both the narrative around AI and the surprising statistical smoothness of reality that AI revealed — to both scare workers into docility and, in some cases, to replace workers with software. If you can do both, the software doesn’t have to be capable of doing the worker’s job, because you can acquire enough market power to produce inferior outputs that the public is still required to consume. That’s the goal of every monopolist: to raise prices and lower quality, which you can only do if you have market power.

This is why competition matters more than a lot of leftists think it does. It disciplines firms because they have to worry about competitors. And the presence of competitors makes it more difficult for firms to solve the collective action problem needed to capture a regulator and to capture a workforce.

I think some leftists believe that the short-lived and extremely contingent large-firm wage premium, where very large firms with a lot of market share paid their workers more out of the excess profits they generated, was a state of nature, and that we need to go back to having really big companies with enough profit to share with us. That’s not true. That was a highly contingent moment in which the hangover from the New Deal coincided with the first days of monopoly — the post–New Deal monopoly surge.

There was enough labor power left to force monopolists to part with some of their ill-gotten booty. But that was never going to be a durable condition, because they were going to accelerate their market power, and that would necessarily dampen labor power. That’s what competition is good for: not as an end unto itself but as a way to shatter the class solidarity of the ruling class.

Conjuring Valuations

David Moscrop

Is the industry in a bubble? You see headlines like “Stock Market Had Its Best Ever Year.” I look at headlines from 1929 and think, I’m not convinced that’s such a good omen.

Cory Doctorow

It is a bubble. I’m not questioning that it’s a bubble. I’m trying to find the ideological and material roots of the bubble. And as a good materialist, I think ideology is downstream of material conditions.

Billionaires don’t like to be thwarted. We can think of billionaire-ism as the pathological state of being angry about being thwarted. We all get called on our bullsh-t from time to time, and sometimes that’s bad because it’s not bullsh-t. But a lot of the time, actually we were wrong, and it’s good that someone was around to call us on our bullsh-t. This is why married people live longer — because there’s someone who can say “No, you’ve got to go to the doctor. No, you have to go and you have to stop eating that.” Being called on your own bullsh-t is good, but billionaires are in a pathological condition of not wanting to be called on theirs.

I also think there’s this other material story about the bubble. Growth stocks have much higher valuations than mature stocks, not because of superstition or a commitment to growth, but because a share is a claim on future earnings. Growing companies have more future earnings than stable ones. Highly valued stocks are also highly liquid, and so they can be used as a substitute for cash. If you want to grow by hiring people or buying other companies, you can use shares to do that if your company’s stock is highly liquid — which is to say, if it’s growing.

Shares are really nice from a firm’s perspective because you can make them endogenously by typing zeros into a spreadsheet, whereas money is something you don’t get to make endogenously without being arrested. You have to get it from banks or investors or creditors or customers. So, firms don’t want to reach the point where markets view them as mature, because they lose that special advantage, and they also lose a ton of their valuation. If you cease to be a growth company, your valuation should have an earnings multiple commensurate with that of a mature company, which is something like 10 percent of the price-to-earnings ratio of a growth company. That means that you become 90 percent poorer the day the market decides you’re not growing anymore.

And I know we say, “Oh, endless growth is the ideology of a tumor,” but it’s not ideological to not want to be 90 percent poorer. That is a purely material phenomenon. On the one hand, you have a narrative need to tell a story about how a company that has undeniably saturated its market — Google has a 90 percent search market share; it has saturated the market — is going to grow into some adjacent sector. It could be one that we can all point to and agree exists like, “Oh, we’re doing Google Plus, so we’re about to become Facebook.” Or it might be a fanciful one that you have to convince people might exist, like Web3, cryptocurrency, the Metaverse, or whatever. But on the other hand, because it doesn’t exist, you can make extremely fanciful claims about its eventual valuation. You’re the expert on it because you just invented it.

And so you have Morgan Stanley saying, “Mark Zuckerberg’s right. The Metaverse is an $8 trillion play.”

Solipsism in the Machine

David Moscrop

Reminds me of NFTs [non-fungible tokens].

Cory Doctorow

NFTs, yeah, which were DOA. AI comes in as the latest version of this. The question isn’t “Is there an AI bubble?” There’s an AI bubble for the same reason that there was a cryptocurrency bubble. The question is “Why is the AI bubble so much bigger?”

On the one hand, it’s so much bigger because the computer science is more real. We should not underestimate how surprising it is to discover the degree of statistical regularity in the universe with this theory-free AI. The fact that you can have a thing that closely resembles a conversation with a thinking person just by guessing words is legitimately remarkable. But it doesn’t tell us anything about consciousness — or, at least, that’s not what I think it’s telling us. What I think it’s telling us is that there’s more statistical regularity than our brains would guess without an instrument, in the same way that our brains guess wrong about all kinds of things.

Your brain guesses how long you’ve been waiting in line: “I’ve been here for an hour.” Then you look at your watch, and it’s only been ten minutes. We persistently underestimate how regular things are, and the only way we can discover that is by having this thing that ingests a lot of data and then makes guesses. It’s cool; it’s interesting. I think philosophy departments can dine out on this for the next fifty years. But it wasn’t worth a trillion dollars. So, that’s part of the reason the bubble is so much bigger: the computer science is real, and it discovered an interesting fact about the universe in a way that none of that other stuff did.

Blockchains are also real; they’re just not as interesting. The number of applications for that kind of slow database rounds to zero. It’s money laundering, gambling, ransomware. That’s sort of it. It’s just not all that valuable, whereas this AI stuff is much more interesting.

But the other piece is the ideological one that I alluded to at the beginning: billionaires are very excited about the idea of a means of swapping money for expertise. I think that’s because being a billionaire also means being a solipsist. You can’t make a billion dollars and hurt people in the way that you must in order to make a billion dollars if you think other people experience hurt the way you do. To aspire to a world in which 99.999 percent of the world receives no moral consideration is to aspire to solipsism. AI is a very solipsistic technology. It’s a technology that sort of implies that everyone else is just a bot. If everyone else can be replaced by a bot, then in some sense, everyone else is a bot.

David Moscrop

Is this version of AI destined to be commodified? AI companies want to rent intelligence back to us by the meter, using content we’ve created. Does data necessarily become a commodity?

Cory Doctorow

The problem is, I think neither of those things are going to work. Selling intelligence as a service requires that the AI companies remain in business. They’re spending a trillion dollars to make $50 billion. Their unit economics just suck.

So, let’s take a little detour into just how sh-tty a business AI is. They’re spending a lot of money to make a little money. That’s not unusual. You spend a lot of money and make a railroad, then make it up by operating the rail carriages over the rails. But there’s a lot of problems with this analogy. In the case of AI, the unit economics — the cost of running the rail carriages over the rails — suck. They lose money on every rail carriage. Every customer they acquire makes them poorer. Every time that customer comes back to use their product, it makes them poorer again.

Every version of AI, every new improvement in AI, loses more money than the previous one. And the depreciation schedule sucks. They’ll tell you that they need to upgrade the hardware every five years. They’re actually upgrading it every two to three years, which is wild, because that’s just accounting fraud. If you’ve got assets on your book that you are treating internally as though they’re worthless, but you’re telling the market and your investors that there’s still 40 percent of their value left, then calling that accounting fraud is doing violence to accounting fraud. It’s just terrible.

And it matters because even five years of $50 billion in annual revenue wouldn’t be enough to recoup a trillion-dollar investment. But their unit economics are not improving. They keep claiming they’re improving. They keep saying, “Well, the cost of inference” — which is the operating expense — “is going down.” It’s not.

Fun With Accounting

David Moscrop

How does the inference model work?

Cory Doctorow

So, you have training and you have inference. Training is when you ingest all the data, and you find its statistical relations. Inference is like when some chud goes to Grok and says, “Make me Sonic the Hedgehog, but with giant tits.” They keep saying inference is getting cheaper, but it’s not. And when you look at the balance sheets, you discover that they’re doing the world’s sweatiest tricks to hide this fact.

Ed Zitron is quite a good and attentive and detail-oriented watcher of AI. Someone leaked him a copy of OpenAI’s balance sheet, and I can’t think of anyone better to leak to because this is the guy who’s going to go over it as closely as you can imagine, with absolute obsessive detail. He reports that the inference line is a small sum of money. But he also reports that if you keep running your finger down the balance sheet, you get to the marketing budget and find that OpenAI claims to spend more on marketing than Coca-Cola.

Coca-Cola has three global marketing agencies with offices in every major city and billboards on every corner. If you want to know whether Coke is really spending money on marketing, you can verify it just by counting billboards on corners and noses in offices. But there is no evidence that OpenAI is spending that kind of money on marketing. Ed’s thesis, which I think is completely true, is that they’re just calling the inference they give away a marketing expense. But that doesn’t make it not an inference expense. That’s like saying “Oh, it’s not a bald spot. It’s a solar panel for a sex machine.” It’s still a bald spot.

These companies don’t have a way to rightsize their balance sheet. And even if they did — and this is where commodification comes in — the switching costs are really, really low. When Anthropic started overtaking OpenAI, everyone just went to ChatGPT and said, “Package up all of my data in a format that’s easy for Claude to understand.” And then, they deleted their OpenAI accounts, opened Claude accounts, and started paying there.

It’s like if you finish building the Canadian Pacific Railway and turn around, and someone has just put down another set of tracks next to your tracks and their trains are 20 percent faster, and you haven’t run a single goddamn carriage over your tracks yet. So, you’ve got to start building your own set of tracks again.

And this is where the open-source, open-weight models come in. The Chinese are taking low-end hardware and open models, and they’re doing something called distillation. That’s when you methodically query Claude and ChatGPT, then use the answers to train your own model without doing the original training. They can’t stop that.

The idea that they’re going to be able to charge rent forever on inference requires, for example, that they cook up a giant scare about how their word-guessing program is about to become God and kill us all because their hacker bot escaped containment. Then they get the US federal government, on the one hand, to make it legal for them to all agree to stop doing R&D — which would otherwise be a flaming antitrust violation — and on the other, to ban all the Chinese models. Right on schedule, they’ve done that. But it’s not going to work.

Copyright Won’t Save Us

David Moscrop

What about the copyright-maximalist response to LLMs training on what we create? Does that further commodify intellectual and creative output?

Cory Doctorow

First of all, the value of your data for training rounds to zero. All we’re doing is creating a crazy version of markets in which affluent people can afford to keep their data out of the model, while poor people accept a pittance — far less than it’s going to hurt them to have their data in the model — and then that data is used in ways that harm them.

The problem with copyright is that it’s not a labor right. Artists are told that it’s their primary mechanism for defending their labor rights, but it’s a property right. They call it intellectual property. Think about it: property rights are fully alienable. If you’re bargaining into a market of five publishers, and four studios, and three labels, and two app companies, and one company that does the ebooks and audiobooks, having more rights just means you have more to bargain away for the same price.

Copyright has expanded for fifty years. Copyright payments have not. The copyrights required as a matter of boilerplate contract by this highly concentrated monopsony that we sell our creative labor to are just 100 percent of all the rights we’ve ever been given. And every time we get a new right, the standard deal expands.

The Hollywood writers were not on strike against Anthropic and OpenAI. They were on strike against the studio bosses, who want to replace Hollywood writers with AI and who are also suing AI companies over copyright. If they win that right, you can kiss goodbye to the socially productive activities that we do that resemble AI training. If you can’t scrape the web anymore, you can’t make a record of what CBS looked like before Bari Weiss took over.

If you can’t publish factual assertions, you can’t publish factual assertions gathered from large corpuses — like how chuddy CBS is now that Bari Weiss is running it, what racial slurs they’re using, how many articles are being written by people of color, or whatever you want to name. That all goes out the window if we create an exclusive right to make transient copies of works for the purpose of analysis, perform analysis on them, and then to publish the results. Copyright does not currently restrict those activities to block AI training; you would have to make them restricted by copyright.

Then bosses in the extremely concentrated creative labor markets would just acquire those rights as a nonnegotiable condition of employment. They would use them to train chatbots and then fire your -ss and replace you with those sh-tty chatbots that hallucinate because they’re just doing statistical extrapolation.

We would then have inferior outputs, worse wages, and a political environment in which it is even harder to hold rich and powerful people to account. If we build out labor rights, privacy rights, human rights, and consumer rights, which are the good rights, and relegate property rights to where they belong — to those things that are so worthless that we can put a price tag on them — then we actually have a chance against AI.

This article originally appeared in Jacobin, a democratic socialist magazine publishing long-form essays and analysis on politics, economics, and culture. Subscribe to the print edition for $20 a year.