This essay comes out of some notes Artie was working on for an upcoming episode of Death Panel podcast. The episode will be coming soon, but we wanted to make part of the argument available in written form first. To hear that episode and support our work visit our Patreon here.
There are a great many reasons to oppose the sudden, rapid expansion of data centers and the ongoing encroachment of AI (which is to say, Large Language Models (LLMs) that have crudely been marketed as a form of “intelligence”) into governance and everyday life. For many who live next to data centers, the noise is unbearable, driving some to habitual use of sleeping pills just to rest. Their operation threatens biodiversity in ecologies already wracked by climate collapse, while themselves being a driver of further ecological catastrophe in an era already marked by drought, rising temperatures, and choking wildfire. LLMs themselves are often used in practice to expand surveillance networks and capacities like Flock’s, and to expand and worsen the carceral archipelago. Basic computing hardware has skyrocketed in price. AI companies are functionally burning books. All to run a couple of programs that no one really asked for and most people outright hate; programs that generate false information and that so regularly drive people who interact with them to suicide that there is an entire wikipedia article cataloguing these deaths. Practically every political ideology imaginable has some issue with the premise of AI eventually putting large amounts of the workforce out of work, with the US’s fascist right even managing to make its slogan echo the rest of their usual racism: they say they want to put “humanity first.”
Note: For much of the rest of this short piece I’ll often refer to data centers and AI—again, actually LLMs—simply as “AI,” because the data centers are the hardware the LLMs run on; one doesn’t exist without the other, and an AI is no more an abstract “intelligence” than “the cloud” was ever any kind of immaterial, ephemeral nowhere-space. (One final aside: people, understandably, hate AI so much that they scrutinize works that use the em dash (—), assuming any text using this punctuation is machine generated. I’ve never willingly used AI and do not plan to. But I love the em dash and will not let AI take it from me. Same goes for parentheticals).
All of these issues with AI are critically important, but I want to address a few overlapping issues here that often go unaddressed or, if not, are frequently inserted into reporting as asides or as background information for which our conclusions should be self-evident. And this is that AI and data centers constitute a kind of double-bind health threat, or a self compounding bodily threat, to all of us. AI as currently constituted operate as a mechanism of social murder. The AI boom is, literally, killing people.
Cancer alley everywhere
The first and most important impact of AI that I want to reinforce here is their status as an acute health threat to the immediate communities surrounding them. One analysis has estimated that by 2028 the combined burden of data centers in the US could be linked to some 600,000 cases of asthma and 1,300 deaths. Study after study has attempted to estimate the massive health burden of individual AI data centers (which is to say, the local effects of even just one data center) on surrounding communities, primarily through annual expected PM2.5 emissions. PM2.5 are particles of air pollution so small that, as the WHO says, they “lodge themselves deep into the lungs and can penetrate into the blood stream.” Exposure to PM2.5 is linked to increased risk of stroke, heart attack, respiratory and cardiovascular disease, asthma, and the over-general catchall we refer to as “premature death.”
According to one study (emphasis added):
There is no known safe level of exposure to PM2.5. Even at concentrations well below current regulatory standards, studies have shown measurable increases in hospitalizations, disease burden, and mortality risk. Although emission-control technologies such as scrubbers and filters can reduce particulate emissions, they cannot fully eliminate them.
I emphasize this because, just as we’ve discussed before with similar issues we’ve seen in activism around safety standards for coal mining, the fact that there is no “safe” way to do these emissions means we should fundamentally question the continuation of this form of economic activity. It’s relevant to note here that AI is itself a driving force in keeping coal power plants running, and therefore maintaining a market for coal mining (which is to say, a market for the premature deaths of workers hired as coal miners).
The conclusions these AI impact analyses come to are, uniformly, stark. Below I’m going to pull some examples, largely sourced from the same team of researchers. All estimates are based on models of the likely health impact of PM2.5 emissions based on the level of PM2.5 the local government has permitted those facilities to produce. Which is another way of saying, the health effects these facilities are being actively, legally allowed, entitled, and encouraged to produce. In technocratic wonk speak, the health “trade-offs” that local officials have approved as an acceptable human cost. More on this in a minute.
One such analysis of a data center in Loudon County, Virginia estimated an average of between 3.4 and 6.5 additional deaths per year in the area surrounding the data center.
Another analysis of a data center in Tucker County, West Virginia, estimated an average of between 1.2 and 2.3 additional deaths per year in the area surrounding the data center.
Another analysis of a proposed data center in Southhaven, Mississippi estimated an average of between 1.9 and 2.8 additional deaths per year in the area surrounding the data center.
Note that each of these deaths per year figures is attached to a single data center.
There are of course plenty of effects beyond deaths—each analysis shows increases in emergency room visits, lost days of work, lost days of schooling, restricted activity days. This fits with impacts we’ve seen from other sources of air pollution that are taken for granted, as if those impacted could be shrugged off as a disposable community—like in Newark, where residents have been forced to take it upon themselves to document how many freight trucks pass through busy intersections in a day (3,000 in one day through just five intersections), and the corresponding health effects; or like the increasingly common effects of wildfire smoke. In this sense, the “promise” of AI is cancer alley, everywhere.
It is not overstating the matter to say that this situation overall feels something like a cautionary children’s fable or a particularly dark nursery rhyme. A town is promised some kind of boon—maybe jobs, economic growth, even when these things rarely actually manifest—and in exchange, one by one people from the town will begin to simply disappear. Never enough at once, or in one year, to raise a crisis; and always seeming to happen at random, as if the product of some unknowable force, even when the machine that’s doing the killing is right there, a big gray slab, humming away and making you lose sleep. Sometimes it takes children, like a fairytale depiction of a witch (with apologies to witches). Sometimes it takes you.
AI defenders will, and do, rationalize these impacts away. If they accept the estimates from the analyses above (which many of them don’t), they will say that one, or three, or six people per year sounds worse than it is because the numbers are expressed as an absolute rather than as a proportion of the population size—that actually, community risk overall is “low.” This is an argument many of us can recognize well from some of the debates in recent years over covid-19. Just because a number can seem small when expressed as mortality per 100,000 population doesn’t make those individual deaths less meaningful; it is people who are dying, not numbers on a spreadsheet. Additionally, we know the direct source of these deaths. It is a decision to not just allow these facilities to be built but to encourage them and stake so much of the economy on the hope that something positive will, one day, come from them.
Defenders will also say things like, “what about the environmental impact of farming???” or compare these deaths to any number of other causes of mortality (once again echoing covid, where a favorite refrain from at least 2021-23 was for commentators to compare covid deaths favorably to car accidents). They will not infrequently point to other industries and say “this industry also pollutes,” or is also bad for the environment, and population health, in some way.
And you know what? Exactly. We live in a society built up into a vast concatenation of death machines of varying intensities. Pointing out that one death machine is not the only death machine does not invalidate criticisms of the first death machine. Defending AI as being not uniquely a cause of death due to industrial byproduct or indifference is, overall, an indictment of our broader economic system, and therefore not a particularly relevant defense of AI. Much of the history of a variety of forms of activism or social unrest can be linked, in one way or another, to people demanding that one death machine in particular be turned off.
Listeners of Death Panel and people who’ve read Beatrice and I’s book Health Communism will know exactly the mechanism of death I’m talking about here: Friedrich Engels’ concept of “social murder,” an idea we have long explored in our political commentary. Social murder is endemic to capitalism and inseparable from how our current political economy forces us to value life in economic terms. It is a term for deaths that occur as a direct result of the various aforementioned death machines whirring away, taking people away at a steady clip in an almost expected, accounted for manner, and largely invisibilized as a result. And a large part of how social murder operates, as our friend Nate Holdren has frequently, eloquently argued, is through getting many of us to accept these deaths and move on. Through, effectively—to try a spin on “manufacturing consent”—manufacturing indifference. Often this happens in popular discourse just by virtue of various stenographers of power trying to stay current with wherever they think the winds are blowing. While not just a media phenomenon, this is for example the general function of any given Atlantic article that says some health threat isn’t a big deal, or the function of any number of Emily Oster articles, or the function of the kinds of headlines that suggest a group of Palestinians have simply died of starvation and absolutely no one, no states, no armies, are to blame. Cedric Robinson called this kind of thing the product of “clever weavers of aristocratic flummery.” He was talking about figures like Plato and Aristophanes, but you can see today’s manufactured indifference as a degraded but continuous line through much of the history of capitalism and settler colonies. The ruling class never met a negative externality it couldn’t portray as a virtue.
In this way, AI (LLMs) are indifference engines. This may be a pun on “difference engine,” but it’s also a very straightforward observation and marks an important shift in the interaction between computing and governance. The roaring hum of the data center is the sound of the indifference engine running, ceaseless and uncaring to the death and suffering it causes.
Organized abandonment, AI and Medicaid
This is also where the second component of AI’s double bind of health impacts comes in. Many of the current uses of LLMs effectively amount to, themselves, manufacturing indifference. I’m not just talking about indifference in the sense of “the computer is reading and responding to your emails for you,” though that is certainly its own social paroxysm. I’m also not just talking about the ongoing, reportedly deleterious effects on scientific research and the effect that may have in the long term on access to basic knowledge or the quality of medical treatment.
What I’m talking about is, in particular, the way governments like the second Trump administration have eagerly embraced AI as a tool of governance.1 And specifically, as a tool of governance, AI offers a mechanism to speed and exacerbate organized abandonment.
There are the surveillance, incarceration, and border policing aspects of the second Trump administration’s broad deployment of LLMs for governance to consider, each of which is its own unique valence of immiseration and death. But I want to consider what’s happening with Medicaid.
Over the last few months, as part of their “War on Fraud”—a whole topic in itself—the Trump administration have paused or withheld roughly $2 billion in Medicaid funds to the states of California and Minnesota. It is no accident that these are states where anti-ICE organizing has been particularly robust, with 2025’s memorable scenes of Waymos on fire in LA and this year’s uprising in the Twin Cities; the Trump administration is clearly only interested in targeting “fraud” to the extent that it can wield this concept as a political cudgel against its various enemies.
This kind of targeted suspension of Medicaid funds is not precedented. To date, the only times it has been done are the few times this Trump administration has done it—and those were all just this year.
Importantly, to justify pausing these funds to states, the Trump administration had to find some way to flag certain Medicaid payments in these particular states as “potentially fraudulent.” The way they chose to do this was through AI. As Robert F Kennedy Jr. said in a July press conference, “We used AI, advanced analytics, and other cutting-edge tools to identify suspicious activity and potential fraud.” (Given this administration’s past history with relying on AI to just make things up, I’m going to just assume that “advanced analytics” and “cutting-edge tools” are mostly embellishments meant to make the process sound more sophisticated, especially since CMS Administrator Dr. Oz has been similarly vague about leaning on AI: Oz wrote last year they would be targeting “fraud” with “artificial intelligence and other cutting-edge technology”).
This might seem like just another example of the Trump administration, and its MAHA contingent in particular, being goofy or careless with facts. This is, after all, a Trump administration, where lies are a little more open and brazen than they often are under others, and where we’re nine—jesus christ—nine years on from the New York Times patting itself on the back for finally deciding to call the man a “liar.”
I think, though, this demonstrates AI as a perfect ideological match for the vicissitudes of the Trump administration. They know they can get away with a lot as long as they have some form of justification for it, and AI happens to be a justification machine. In the case of Medicaid, this is particularly insidious. The Trump administration isn’t setting out to outright prove that fraud is occurring in any given Medicaid provider to withhold funds; they have been, repeatedly, very clear that they have used AI (and whatever other putative “cutting edge” technology they’ve handwaved about) to generate a list of providers that could, maybe, have some fraud or overpayment going on. They are asking the machine to guess. They then use that list as a justification to withhold critical funds to states where its agenda has met resistance.
The key here is that for these “paused” Medicaid payments to not become de facto program cuts, states have to then take on the administrative task of gathering proof that those payments are not fraud. (This aligns perfectly with all of the Trump administration’s other schemes to take more unilateral control of funding through the executive branch and turn fiscal federalism into, as the political scientist Phil Rocco says, “a more or less explicit system of political patronage”). This is effectively additional make-work for states that already do a fair amount of internal auditing on their own, but can be directed to dump hours and hours of administrative time (or the added expense of contractor hours) into proving everything is above board just so that regular payments can resume. This puts the federal government in a position where it can effectively generate additional administrative burdens for states at will, with minimal effort. States, presumably having a higher burden of proof than the federal government’s simple act of “hey, our liar machine says this provider might be fake,” will be stuck putting time and money into trying to re-secure the funding they’re supposed to be guaranteed, instead of spending that time and money on administering their Medicaid program and getting people healthcare.
The ultimate effect that will have is straightforward, and fits with what we know about the rest of Trump’s agenda on healthcare, part of which is to kick as many people off Medicaid as possible. When the federal government reduces Medicaid payments to states, states often adjust eligibility criteria to make it harder to get or stay on Medicaid (for states, less people on the program means less expense, and less chance of running through their budget since they can’t deficit spend). When states make it harder to get or stay on Medicaid, more people get kicked off the program. Copious research agrees that people losing healthcare coverage, even temporarily, is linked to adverse health outcomes, including premature death.
In essence, by using AI to flood the zone on state bureaucratic systems while actively withholding billions in funding, the Trump administration is able to more or less guarantee program cuts and individual death and immiseration without having to go through the time- and political capital-consuming process of legislating those cuts (which, to be fair, they have also done in the form of work requirements). It should come as no surprise that the Trump administration is simultaneously pursuing cuts to research concerning the health impacts of changes in public policy—the very kind of research that generates estimates like the annual death figures discussed above.
Crucially, AI also allows the Trump administration to in part bypass the otherwise difficult step of whipping up a coalition of patsies to do all the work of burying states in paperwork; one Trump official recently said they would not have the “time or manpower” to do this otherwise. (“In part” is the important qualifier there, because there is still only so much that AI can do as a tool of governance—it can identify purported “dissidents,” for example, but for now you still have to have a willing, human police force to do an arrest). This is important because traditionally fascism requires the loyalty or at least acquiescence of a large amount of people ready to carry out aspects of the agenda, a task that some have argued Trump has actually been very bad at. The worry that we should all hope does not fully manifest is that, in AI, the Trump administration finds it doesn’t need a bunch of goons who are indifferent to the suffering of others or who actively revel in it; the machine is already built to do exactly that.
Final thoughts
As many have pointed out, the mass infrastructure buildout that is currently being sustained in order to build these death machines is proof positive that any of the big ideas for reorganizing society or the relations of production brought by communists, anarchists, or socialists—from Medicare for All to abolishing property and hierarchy—are eminently attainable; that demands are not “too big,” but instead just don’t align with the interests of those in power. When the forces of capital and the capitalist state want to engage in revolutionary acts of economic and social disruption, they will do so at any cost, and they more often than not get away with paying that cost in the form of our lives, our blood, our debilitation. In the case of AI, this may carry huge implications for fascist governance.
Further, when politicians, elected officials, and those otherwise in power support data center construction (and continued operation) they are endorsing, at once, a mechanism of increased death in their communities and the very operational state capacities the federal government is currently using to extort subnational governments, build lists of dissidents for a new Red Scare, and round up migrants for internment, deportation, or execution. When Democrats talk about reformist maneuvers to make data centers act more ‘responsibly’ (as PA Governor Josh Shapiro has) they are talking about, at most, tweaking the dial on the indifference engine to try to negotiate just slightly fewer deaths but not eliminate them entirely. When Democrats talk about reforms to make AI companies “pay for disruptions caused to communities” (as Senator Ron Wyden has) they are talking about collecting blood money (and at that, blood money that more than likely will never come).
AI’s defenders likely believe they can weather the storm of public opinion, as capital is so used to doing, but we should expect and intend opposition to grow much more extreme as the negative impacts of AI become more clear to more of us. The fact that massive opposition has emerged to these indifference engines should not surprise us, but it should also motivate us to focus that energy in liberatory directions. As a novel productive capacity growing at an alarming rate despite immense public disdain for the technology, part of the lesson here is how we are all forced over and over again to accept new conditions of death and immiseration as they emerge until such a time as that immiseration can be made to seemingly fade into the background noise of polite society. While our society may contain a vast concatenation of death machines of varying intensities, that does not mean we should oblige another death machine being turned on. We can dream instead of a world where the datacenters are torn down and forests grow in their wake.
