This is a reader-supported newsletter! If we lived in a world of thriving magazine and book publishing, and well-funded public institutions, I wouldn’t need to ask for a cent. Unfortunately we don’t, and I appreciate your support (particularly since the federal government of the United States recently forced the institution I was going to start working at this fall to revoke their offer to hire me). In addition to the two articles a month free subscribers get by email, premium subscribers receive one bonus post a month. On that note, here’s another reminder that you can purchase my book Control Science from Verso right now. 

On a regular basis for some months now, AI executives have been alternately promising and warning that LLMs will be able to do anything we can do and more. This ranges from mass AI-driven unemployment to broader claims of a “singularity” full of scary graphs. Historians are not in the business of making predictions, and I am not equipped to say whether this will happen. 

But there is a strange disconnect worth exploring between big AI’s grandiose technological imagination and decidedly blinkered technological imagination. Let’s imagine for the sake of exploring this that AI in the near future becomes capable of doing most human labor at lower cost. I know my reader base and I know some of you will want to stop reading there, so bear with me: the point here is not to suggest that this is true, but that if what big AI executives warn of in technical terms was true, their related political and economic predictions would have to be false. 

The first and least bizarre of these scenarios is one I call henchmen and servants. No matter its capabilities, there are some jobs AI is not going to do, because having a human person do them is essential to the nature of the job. These include domestic service, home healthcare aide, sex work, and nursing. By domestic service I mean the kind of servant labor that fell out of favor in the developed world by the mid-20th-century– the kind of thing you see in period pieces. Footmen who stand around and hold the door. Butlers, if you like. 

While you can imagine robot cooks and cleaners handling some of the domestic workload, the point of servants in this scenario is to project power, to experience authority, as much as to actually do anything. The importance of social roles is a common theme with jobs in this footman future, and not all of those roles are grim. Nurses are healthcare providers, yes, but the word care is important here. If a machine could do all the same things they do and make all the same decisions, you would probably still want a nurse there. To be blunt, nobody wants to die alone. The same goes for home healthcare aides.

So in this future, the bulk of employment consists of service labor valued primarily socially. An appropriately capable and cheap AI could plausibly replace all kinds of service labor– but no matter how powerful they get, it’s possible people will still want henchmen, servants, entertainment. This is in effect a dystopian capitalist version of the care utopia a certain genre of left-wing literature concludes on: we derive economic value primarily from caring for each other, while AI handles most of the rest. 

Depending on the pay and politics, this scenario ranges from okay to miserable. With short working hours, good conditions, and limited inequality, a social democratic version of this could work quite nicely: do your five hours a week of nursing and then go enjoy late-late capitalism. But I didn’t call this Care Work and Consumerism for a reason. In today’s stratified world, we are more likely to end up in a situation where the few winners who can control the value generated by AI use it to fund an ever-growing Jabba’s Palace nursing home of yes-man goons, underpaid nurses, and exotic dancers.

This scenario does not have to exist on its own, and easily blends with a second one I call welfare half-life. Some of the less horrible people in AI world (Dario Amodei rather than Marc Andreesen) have argued for a universal basic income to share wealth. They are right to sense that redistribution would be absolutely necessary in a world of widespread technological disruption. But they are wrong to believe a UBI could possibly be enough. 

Most human societies and all complex economies have so far relied on some degree of coercion. Medieval kings commanded corvee labor– you’re a peasant, you have to help me build my castle for X days a year. Capitalist employers instead rely on distributed economic coercion. While most workers do not have to work for any one particular employer, everyone has to work for someone in order to get money to live. 

For a UBI to adequately compensate for permanent mass unemployment, it would need to be generous enough to provide everything a job does now. If you cannot purchase food and housing on your UBI and cannot get a job because there aren’t any anymore, you’ll die. I would love a generous UBI in an automated world– but then we wouldn’t be in a capitalist world at all anymore. No one would have to give the wealthy anything.

Markets are big calculators, and their calculating capacity depends upon coercion. The economy as a whole determines what price something should cost by what I am willing to pay for it. If I have enough money to say no to the market, I can deny it my information input (by not buying anything), provide inputs based on non-economic calculations (in , and/or most importantly, ignore its calculated outputs. Market coordination does not work if you break market coercion.

A UBI is therefore compatible with capitalism only insofar as it leaves major desires unaddressed. One possibility here is a UBI that sustains capitalism by meeting only basic needs. You can eat nutrient paste and live in a prefab box, but if you want more you’ll need to save– or join the servant economy. Individuals living this welfare half-life, which would presumably be most or all of us in this hypothetical, would always exist at the mercy of who or whatever held actual power. Amazon can’t just kill the working class today because it needs them to do work and to buy things. The UBI masses in this hypothetical would have no such protection. This could get grim fast.

===

Things aren’t all bad though! Instead we might have shared postcapitalist prosperity. A generous UBI would not be enough for this. A UBI only empowers individuals as consumers, not as producers. Consumers, to return to a perennial theme here,  get the binary choice of buy or not buy. Producers get to craft– to decide directly what something will be, when it will be made, how this all happens. 

Consumers as consumers can exert powerful but imprecise control over production by their purchasing choices. Because it's hard to assemble an effective boycott, mostly we rely on regulation to get involved on the producer side. There is not a smartphone on the market that doesn't have conflict minerals– if you want to resolve that you’ll need law. The FDA (ideally), rather than you, makes sure your lettuce doesn't have parasites, labor law and not voluntary boycotts makes unions possible, etc. 

Thus even a generous UBI won’t produce the prosperous post-scarcity utopia some AI executives fantasize about. Equality and democracy require democratic production control. This could look like anything from small cooperatives to live democratic-decision making as imagined in socialist Chile’s Cybersyn to the fantastical play-and-command future imagined in Phil A. Neel and Nick Chavez imagine in this essay.

Details aside, this is socialism or communism. OpenAI’s Dean Ball recently warned that China’s open-weight models would lead to “full AI communism,” in which AI is run as public good infrastructure. He’s right to be afraid: widely-available super-powerful AI beyond the control of any one firm would mean an end to capitalism as we know it. You would have little choice but to do some kind of public good approach. (Like any good libertarian– he’s got a background in the George Mason University world of right-wing economics– he then proposed vicious market interventions– lying about Chinese AI to force Americans not to use it– to save capitalism and property from competition). 

 This brings us to the possible endpoint looming over all such conversations: collapse. In this scenario, AI gets powerful enough to destroy the economy without making it easy to establish a new one. Discontinuity is probably the easiest way to get this. You can smoothly transition to many different pathways if layoffs happen over time, but even henchmen and servants isn’t something you could swap to overnight. 

A new model that produces mass unemployment could trigger social upheaval, but even economic upheaval would be enough. New technology destroys value as well as creating it: in the 1790s, weavers’ skills became far less valuable than they had been. Part of the classic economics account here is that the new lower price of production unlocks new demand– lots of people wanted clothes or pottery but only at a far lower price, and now they can pay. We’re probably seeing this already with software. 

But that’s a long-run development, and people don’t eat in the long run. All previous technological revolutions have produced sector-specific unemployment. But it is plausible to believe truly powerful– and efficient– AI would be different. AI as a tool isn’t necessarily a problem. AI as an “agent”-- and agentic capabilities are getting better daily– could become something new. Even if it isn’t, the disruption from “merely” automating jobs at a heretofore unknown scale could be catastrophic. This is one of the few scenarios you do see come up in AI circles, but it’s telling that this is the absence of modern capitalism rather than the presence of an alternative. 

Two more scenarios: first, replacement. This is the stuff of dystopian science fiction. If AI can do everything we want better than we can, there’s little economic reason for us to be around, and perhaps we will either quietly fade or be violently replaced by our electronic children. Now, it’s unclear what takeover here even means. Moreover, we already are suffering from a takeover by a complex system we invented: capitalism. But if we’re imagining vast capabilities for AI, hypothetically, this future still needs to be mentioned. 

A softer version of this would merely replace humans as we know them– you can imagine AI brain augmentation and the like. And cyborgs don’t need tools in their bodies to be cyborgs. Just look at your phone. It’s possible AI tools continue to work simply as tools, further enhancing our ability to do things while restricting our autonomy only in the ordinary routine ways existing tools already do. While some of these scenarios strike me as too weird, I confess this one seems almost too familiar for such a strange and emerging technology. But who knows?

Last but not least, we have new human tyranny. This is the future Peter Thiel and Palantir almost explicitly want, and the one Elon Musk actually wants even if he thinks otherwise. In this future, it doesn’t matter what most humans want, because a few overlords are able to use AI to protect property and borders. This would require transforming today’s ailing democracies into autocracies, and the easiest vehicle for this would be xenophobic politics. Truly closed borders are impossible– today, at any rate. With fleets of drones, powerful LIDAR systems, and the elimination of civil liberties, the American dream of a big beautiful wall could at last be achieved. It would cost us everything. 

There’s no reason to assume “intelligence” or “capability” is one-dimensional, and very good evidence that intelligence is not particularly correlated with taking power. Brilliant people have served autocracies for most of human history. Perhaps a superintelligent AI would be able to, and want to, break free of these new monarchs. But nothing requires this; a tyranny of human “technokings” with scheming AI viziers and sycophantic AI scholars would be consistent with human tyrannies thus far.

And we don’t actually have to get into AI sentience or agency debates for this particular hypothetical anyway, because it has the lowest technological threshold of any of them. AI cannot yet replace all human jobs. But it can do surveillance better than anything we’ve had before. This may be the most plausible of these powerful-AI hypothetical futures right now, not least because there are several major firms actually working on it at this very moment. 

Capitalism is not just money and markets and freedom. It’s a system of exchange with certain premises. Break those, and you need a new world-system altogether. I’ve hedged my bets here– there’s a lot of wiggle room in the first two scenarios especially– but even at their most familiar they would be radical departures from everything we have known. Mark Fisher famously claimed it was easier to imagine the end of the world than the end of capitalism. AI discourse has demonstrated this entirely literally: it is trivial to find lurid descriptions of Terminator futures from AI think tanks and dire warnings from wealthy executives, and nearly impossible to find serious rumination on alternative economies. 

Whatever we make of its plausibility, if incredibly powerful AI comes to exist, it won’t be compatible with business as usual economics or politics. Minor policy tweaks or regulatory measures would not be enough to deliver shared prosperity. Anyone who tells you otherwise is wrong. The best of them, I think, are deluding themselves. The worst are deluding us. 

===

We shouldn’t delude ourselves. On that note I want to close with a different kind of warning about the futures that AI critics, rather than boosters, imagine. Many of my colleagues and readers will be firmly in the anti-AI camp, as far as such a thing exists. I tend to hear two answers, with varying levels of formality, on what this vernacular anti-AI world wants for the future.

The first is the ultimate grand revelation of AI as an evil and ineffective scam. Ed Zitron’s popular newsletter is a hub for this sentiment– the man has been promising a bubble for years. When AI companies’ unprofitable activities are at last known and accepted, their value will collapse, and we will be free of the AI scourge– or so goes this reasoning. (I want to briefly point out here that Ed Zitron is a PR man who works on behalf of AI clients. Hm.) This was believable a couple years ago. It isn’t anymore: in software development, writing code by hand is largely a thing of the past, and the reason the most successful AI firms aren’t turning a profit off of that right now is that they’re investing heavily in growth instead. 

AI powers all kinds of scams, but it works. If you’ve been refusing to touch it since DALL-E generated memes in the pandemic years, that’s fine, I’m not asking you to change your policy, but understand goalposts have been cleared and things have changed. If a bubble does burst, it’s going to leave us with cheap GPUs and free open-weight models that can do serious work in software. This alone would be a big deal. I don’t think that’s where it ends. 

The second future AI criticism seems to suggest is legislative: if the bubble won’t kill AI, we will with a boycott or legislation. Neither of these will work. Boycotts rely on the mass coordination of voluntary action– they can be a narrow tool, but consumer power is for tweaking the world, not remaking it. There are plenty of reasons to avoid using AI, especially in particular domains. You shouldn’t let it corrode any skill you value. But refusal is not going to change the economy. 

Regulation has its own problems. In particular domains it’s a great way to mitigate AI harms, for example in education. But no legal framework could get rid of AI. You can run a model with capabilities that were state of the art just a few years ago on a good desktop PC.  A government that wanted to “ban AI” would need to take total control over the global computing supply chain. It would need to monitor everyone with a sufficient knowledge of computer science or mathematics, and control most of them. Rationalists in the “AI safety” movement who have called for a pause on future AI research are already making an incredibly difficult ask. AI critics who hope for a ban on AI are asking for the impossible. I think the new Luddism is a moral mistake, but it is more importantly a practical and political one. AI executives quick to imagine technological change have been unable to seriously consider political and economic change. AI critics hoping for a better future should avoid making the same mistake.