Beneath the abstractions of petroleum economics and energy futures lies a blunt material fact: oil is fossilized organic matter. Ancient plants and animals compressed over geological time, their decomposed bodies transformed into hydrocarbons. When we extract and burn oil, we are drawing on the concentrated remains of organisms that lived hundreds of millions of years ago: dead bodies, preserved and accumulated, now made to power the systems of production in the present. Oil is undead in the most literal sense: past life reanimated, put back to work.

In my forthcoming book, Undead Labor: Capitalism’s Zombie Worker (Minnesota University Press, 2026), I examine how capitalism produces figures of laboring bodies that persist beyond the normal boundaries of life and death–bodies from which value continues to be extracted long after they should, by any ordinary measure, have ceased to produce. The zombie, emerging from Haitian plantation culture as an allegory of the slave, serves as the book’s paradigmatic figure: a corpse raised for the purpose of working, severed from consciousness yet continuing to toil. But the zombie is not the only figure of undead labor I examine. The book also traces this logic through the HeLa cell line, the “immortal” cells harvested from Henrietta Lacks in 1951 that continue to labor in laboratories worldwide, generating scientific knowledge and commercial value decades after her death, as well as through the Blob from the 1958 horror film, and through the contemporary proliferation of artificial intelligence. Each of these figures reveals something about capitalism’s drive to extract value from bodies in excess of what they are structurally able to produce. The following discussion of data and oil runs parallel to one that will appear in my book, exploring the phrase “data is the new oil.” In Undead Labor, I consider the way that data functions as both representation and remainder, exploring the ways that the data of dead users might be used to create “digital zombies.” While this essay draws on and extends my analysis of the phrase, it takes a new direction and doing so allows for new insights into undead labor to come to light around temporal violence and exhaustion.

In 2006, the British mathematician Clive Humby coined a phrase that has since become ubiquitous in tech industry discourse: “data is the new oil.” Humby intended an analogy about value and refinement; raw data, like crude oil, must be processed to become useful. But the metaphor carries more than he intended, inadvertently showing how both oil and data bear the marks of what I write about as undeadness and undead labor. To say that data is the new oil is to say that data is the accumulated traces of past human activity, expression, labor, or life, that has been translated into data, in other words, preserved and made available for extraction. It is to suggest that the tech industry’s relationship to data is structurally similar to industrial capitalism’s relationship to fossil fuels: both involve drawing value from traces of the dead.

For nearly two decades, the comparison between data and oil seemed aspirational. Where oil was finite, toxic, destructive, data appeared clean and unlimited–a resource that could be endlessly generated through new tactics of datafication. But in October 2025, the chief data officer of Goldman Sachs, Neema Raphael, made a startling announcement: “We’ve already run out of data.” The large language models and generative AI systems that have consumed the tech industry’s attention, he explained, have been trained on so much human-generated content that the supply is now exhausted. The internet has been sucked dry. His proposed solution was something the industry calls “synthetic data”: AI-generated content used to train future AI. Yet Raphael warned that this risked a “creative plateau”–a point at which the systems, feeding only on their own outputs, would cease to produce anything new.

There is something distinctly surreal about this announcement: namely, that the sum of human expression scraped from the internet–ranging from classic works of literature and philosophy to academic papers and corporate white papers, to Reddit threads and YouTube comments–could be a limited resource. Raw material that could be consumed, depleted, and used up. Stranger still, the proposed solution was for the very systems that are trained on this data to generate their own food, to sustain themselves on their own outputs. The machine learning industry has become an ouroboros, the serpent devouring its own tail (Figure 1). I want to propose “capitalist surrealism” as a name for this condition, paraphrasing Mark Fisher’s famous notion of “capitalist realism,” which he describes as “the widespread sense that not only is capitalism the only viable political and economic system, but also that it is now impossible to even imagine a coherent alternative to it.” In contrast to capitalist realism, capitalist surrealism describes a world in which capitalism, now the globally ascendant political economy, begins to turn in on itself, populated by cryptocurrency scams, machine hallucinations, and “AI slop.” Capitalist surrealism names our experience when the aesthetics of power no longer need to cover its true aims under the mask of good intentions.

Figure 1. Ouroboros, from a Byzantine Greek alchemical manuscript, 15th century. Biblioteca Nazionale Marciana, Venice (Codex Marcianus graecus 299, fol. 188v). The serpent, rendered in red and green to signify the union of opposites, consumes its own tail–a symbol of cyclical transformation in alchemical practice. The Greek text within the circle contains alchemical formulas and precepts. In the context of synthetic data, the ouroboros takes on a different valence: not eternal renewal but terminal self-consumption, a system feeding on its own outputs in a desperate attempt to sustain itself.

A famous passage from the first volume of Marx’s Capital helps to illuminate the stakes of this ouroborian logic, “Capital is dead labor that acts like a vampire: it comes to life only when it drinks living labor, and the more living labor it drinks, the more it comes to life.” For Marx, “dead labor” names the accumulated past labor embodied in the means of production, its machinery, tools, and raw materials, which can only be activated and produce value through the consumption of living labor in the present. Capital’s vampiric character lies in this dependency: it feeds on the living to animate itself. I develop the notion of undead labor as a third term in the binary of living and dead labor that Marx describes in his critique of political economy and go further to suggest that capital, in its never-ending drive to produce value, creates conditions of dispossession that persist beyond the boundaries of life and death. Where Marx’s dead labor is inert until animated by the living, undead labor names a condition in which labor itself continues, extracted and reanimated, without the living laborer’s presence or consent.

The extraction of fossil fuels involves a particular form of temporal violence: drawing down approximately 300 million years of accumulated biomass in roughly two centuries. The modern consumption of oil is the liquidation of deep time, the conversion of the geological past into the geopolitical present, not to mention a future catastrophe in the making. This extractivist impulse always leaves something behind: the tailings, the waste rock and residue that accumulate at mining sites, the toxic excess that could not be recuperated into productive power. Data extraction operates on a compressed but analogous temporal logic. The “high quality” content built up over decades–blogs, forums, articles, images, and conversations that constitute the internet as an archive–is being consumed in a matter of years by ever-larger machine learning systems. But where oil’s source organisms are ancient, data’s sources are recent, identifiable, us. The violence of extraction is not displaced into deep time but operating on the immediate past, on the still-warm traces of human expression.

This temporal proximity raises a question the industry has barely begun to confront: what is the status of data generated and left behind by those who have died? Thousands of users die every day across major platforms, leaving behind archives of posts, messages, images, and interactions. This data does not disappear; it persists on proprietary servers, a potentially valuable resource. The data of the dead represents a starkly literal form of undead labor: the traces and echoes of human activity continuing to generate value (used to train models, to refine algorithms, to feed the vampire) long after the humans who produced them have ceased to exist. If corpus and corpse share an etymology, then the training corpus is also, in a sense, a grave: the accumulated remains of human expression, exhumed and reanimated for purposes their authors never intended and could not have foreseen.

It is in this context that synthetic data emerges as the industry’s proposed solution to exhaustion. The term “ouroboros” derives from the ancient Greek oura, meaning “tail,” and -boros, meaning “eating”–the word translates quite literally as “tail-eating.” This etymology resonates beyond the obvious image of self-consumption. In the language of mining and other extractive industries, “tailings” names the waste materials left behind after ore has been processed: the toxic residue that accumulates at the edge of the mine, the remainder that cannot be profitably used. The ouroboros of synthetic data suggests a system that has begun to feed upon its own tailings, consuming the residue of previous extraction in a desperate attempt to sustain itself.

The trouble with synthetic data is that the temporal loop of production and consumption tightens into a closed circuit; it takes the final turn towards simulacrum, in which any reference to the real finally falls away and is lost to endless recursion. Researchers have begun documenting “model collapse,” or the degradation that occurs when AI trains on AI-generated content over successive generations (Figure 2). The outputs become increasingly homogeneous, increasingly detached from reality as an accurate expression of a world outside of representation. Difference begins to disappear as sameness creeps in, building a system of reference that is increasingly self-referential. The “creative plateau” that Raphael warned of is an inadvertent admission of the unsustainability of the system. Marx’s vampiric capital cannot live on dead blood alone; the snake is eating its own tail. This is capitalist surrealism at its most literal: a system so committed to extraction that, having exhausted its external resources, it turns inward, feeding on its own output, hallucinating sustenance and attempting to feed on a simulacrum.

Figure 2. Model collapse in Variational Autoencoders (VAEs) trained on the MNIST dataset. When each successive generation of AI is trained exclusively on data generated by the previous generation–without access to original human-produced data–the outputs progressively degrade. The original model (a) produces recognizable, distinct handwritten digits. By generation 5 (b), the digits begin to blur and lose definition. By generation 10 (c), individual digit classes become difficult to distinguish. By generation 20 (d), the outputs have collapsed into homogeneous, undifferentiated forms—a visual representation of what the authors term “model collapse.” Source: Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, and Ross Anderson, “The Curse of Recursion: Training on Generated Data Makes Models Forget,” arXiv:2305.17493v3 (April 14, 2024), fig. 9.

What the ouroboros of synthetic data reveals is the irreducibility of living labor to the systems that feed upon it. The current methods of training and developing AI require “fresh” data, the novelty born out of lived experience, just as capital needs living labor to animate its dead machinery. No amount of technical ingenuity can escape this dependency, only defer and disguise it. Synthetic data is the hallucination of extraction without a source, of value without labor: it is the dream of a capitalism that has finally freed itself from its dependence on the living–a dream that began with the zombie and whose most recent iteration has taken the form of AI. But the dream cannot sustain itself. What the discourse of data exhaustion reveals is that AI’s apparently autonomous intelligence is stolen time; the accumulated hours of human thought, activity, and expression that have been extracted and transformed into a resource that appears to belong to no one. Somewhere beneath the surreal surface, the question of labor, of who works, for whom, under what conditions, and at what cost, persists, undead, waiting to be asked again.

References

Bering-Porter, David. Undead Labor: Capitalism’s Zombie Worker. Minneapolis: University of Minnesota Press, 2026. https://www.upress.umn.edu/9781517917852/undead-labor/.

Fisher, Mark. Capitalist Realism: Is There No Alternative? Winchester: Zero Books, 2009.

Goldman Sachs. “AI Exchanges: The Role of Data.” Goldman Sachs Exchanges (podcast). With Neema Raphael, George Lee, and Allison Nathan. September 30, 2025. https://www.goldmansachs.com/insights/goldman-sachs-exchanges/ai-exchanges-the-role-of-data.

Marx, Karl. Capital: A Critique of Political Economy. Translated by Ben Fowkes. Vol. 1. London: Penguin Books, 1976.

Shumailov, Ilia, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, and Ross Anderson. “The Curse of Recursion: Training on Generated Data Makes Models Forget.” arXiv:2305.17493v3, April 14, 2024. https://arxiv.org/abs/2305.17493.

Yeaworth, Irvin S., Jr., director. The Blob. Tonylyn Productions, Paramount Pictures, 1958.