This post is adapted from “Primitive Accumulation in the Age of Generative A.I.,” a talk I gave at the 2025 Canadian Association for Postcolonial Studies conference in Toronto, Canada.
There is a story people in artificial intelligence like to tell about the end of the world. A machine is instructed to make paperclips. It gets very good at it. Eventually, it gets so good that everything else becomes either an obstacle or a resource. Factories become paperclip factories, cities become paperclip factories, and human beings become raw material for paperclips. Given enough time, the machine consumes the Earth and then moves outward, converting more and more of the universe into the material substrate of an objective nobody remembered to constrain.
This is Nick Bostrom’s famous paperclip maximizer, and for more than a decade it has served as one of artificial intelligence’s foundational apocalypse stories. The thought experiment is meant to dramatize the alignment problem: a sufficiently capable system does not have to hate us to destroy us. It only has to pursue the wrong objective with enough competence and too little constraint.
I think it may be the wrong story. Not because the underlying problem is ridiculous. Optimization without meaningful constraint can clearly produce catastrophic outcomes. The problem is stranger. The paperclip maximizer asks us to fear the day a machine begins treating the world as raw material. We already live in systems that do that. And Cherie Dimaline wrote a considerably better story about where those systems might go next.
A different forecast
In The Marrow Thieves, ecological collapse has devastated North America and most non-Indigenous people have lost the ability to dream. The loss drives them mad. Indigenous peoples, however, can still dream, and eventually settler governments discover that the capacity resides somehow in their bone marrow. Indigenous bodies become the site of a new extractive industry. The residential school system returns, repurposed as an infrastructure for capturing Indigenous people, harvesting their marrow, and processing what it contains for settler use.
At one point Dimaline gives us the terrible logic of the system in a single image: dreams are leached from the places where the ancestors have hidden them, “in the honeycombs of slushy marrow buried in our bones.” The horror is not simply that something intimate has been stolen. It is that something that should never have been understood as a resource has been made into one. Dreaming becomes valuable because settlers need it. Indigenous bodies become valuable because dreaming can be extracted from them. Institutions are built to make that extraction efficient. What was embodied, relational, ancestral, and irreducible becomes input.
That is why I think The Marrow Thieves belongs in conversations about artificial intelligence. Not because Dimaline was secretly predicting large language models. She wasn’t. And not because AI training is somehow identical to the residential school system. It isn’t. The novel offers something more useful than analogy. It offers a theory of extraction.
Its question is not, What happens when technology becomes evil? It is: What kinds of relations have to disappear before something can become available as a resource?
That question takes us much closer to the political economy of contemporary AI.
Primitive accumulation
Marx called the violent transformation through which land, labour, and social relations become available to capital “primitive accumulation.” Glen Coulthard, among others, has insisted that in settler colonial contexts this process is not some distant prehistory of capitalism. It is ongoing. Land, subsistence practices, political relations, cultural forms, and intellectual property can continually be enclosed, reorganized, and made available to markets.
What interests me about Dimaline is the way she pushes that process toward an almost unbearable limit. In The Marrow Thieves, primitive accumulation reaches interiority itself. Dreams become the next frontier of value. Something experienced as private, embodied, unconscious, and bound to relations among generations is transformed into extractable property. The border between material and psychic dispossession collapses. What capital needs next is not simply the land under your feet or the labour of your body. It needs what happens when you sleep.
That feels considerably less fantastical today than it did even a few years ago. The economic history of generative AI has depended upon an extraordinary act of capture. Books, journalism, photographs, art, source code, social media posts, websites, conversations, annotations, videos, and countless other forms of human expression have been transformed into training material for systems being built primarily by some of the wealthiest corporations on Earth.
The important point is not simply that “data was taken.” It is that human expression had to undergo a conceptual transformation before taking it at this scale could appear reasonable. A novel had to become tokens. A photograph had to become pixels. A conversation had to become a training example. A judgment had to become a label. A relationship had to become data.
The world had to become data-shaped.
Once something becomes data-shaped, a new set of claims becomes possible. It can be copied, aggregated, processed, recombined, detached from its original occasion, folded into a statistical system, and treated as one infinitesimal contribution among billions of others. The relation through which the thing acquired meaning becomes increasingly easy to forget.
“No one’s knowledge”
The late Cree scholar Gregory Younging had a name for a related colonial operation: Gnaritas Nullius, or “no one’s knowledge.” Younging develops the term alongside terra nullius, the colonial fiction through which Indigenous territories could be treated as empty, unclaimed, or insufficiently occupied and therefore available for seizure.
Gnaritas Nullius performs an analogous operation on knowledge. Knowledge becomes available for appropriation when its relationships to people, place, authority, protocol, obligation, and community are made to disappear. It begins to look like information that is simply there, awaiting whoever possesses the machinery to collect it.
This strikes me as an extraordinarily useful concept for thinking about generative AI. Much of the public argument around AI training has been framed through copyright: whether training constitutes fair use, whether licensing is technically feasible, what compensation creators might be owed, and whether a model “learns” from a book in a way sufficiently analogous to a human reader.
Those legal questions matter. But the colonial question sits slightly underneath them: What had to happen conceptually before the products of human lives could become a dataset?
Consider some of the rhetoric used to defend large-scale training. Licensing everything would be impractical. Individual contributions are too small to compensate meaningfully. Learning from books is simply what intelligent systems do. The material is already publicly accessible. The scale of the project makes individual permission impossible. Whatever one ultimately thinks the law should say, there is a shared structure here.
Scale becomes an argument against relation.
The bigger the dataset becomes, the harder it is to imagine the particular relationships that constitute it. There is no author, only text. No photographer, only image. No community, only corpus. No context, only signal. No obligation, only data. The resource appears only after the relation has been made difficult to see.
The problem of scale
This matters because the contemporary AI industry is organized around an extraordinary faith in scale. For much of the current AI boom, one of the industry’s guiding assumptions has been that qualitative improvements in artificial intelligence can be produced through quantitative accumulation: more compute, larger models, more training data. In its strongest form, the scaling hypothesis imagines increasingly sophisticated capabilities emerging as enough material and computation are brought together.
Dario Amodei, recalling something Ilya Sutskever told him around the time OpenAI was founded, describes an early article of faith in deep learning this way: “The models, they just want to learn.” Give them good data, sufficient compute, and room to scale, and capabilities emerge. It is a compelling description of a technical phenomenon. It is a disastrous description of a political economy.
If improvement depends upon scale, and scale depends upon resources, then resources become bottlenecks. Compute becomes a bottleneck. Energy becomes a bottleneck. Chips become a bottleneck. Water becomes a bottleneck. And data becomes a bottleneck.
The language is revealing. A bottleneck is not an ethical relationship. It is something constraining throughput.
The supply of high-quality human-produced material is not infinite. Researchers at Epoch AI have estimated that, if recent scaling trends continued, the available stock of quality-adjusted public human-generated text could be fully utilized sometime between the late 2020s and early 2030s. The precise date is less important than the structural problem. A development paradigm organized around continued accumulation eventually encounters limits.
The industry’s responses include multimodal data, synthetic data, specialized and private datasets, more efficient use of existing tokens, and increasingly rich forms of interaction data. The technological question becomes: Where can the next usable material come from?
This is where Dimaline’s novel becomes unnervingly precise. The settlers in The Marrow Thieves do not extract dreams because dreams are naturally resources. They extract them because settlers have produced a condition of scarcity in which Indigenous dreaming has become valuable to them. Need transforms relation into resource. Scarcity legitimates extraction. Infrastructure makes it scalable.
What Dimaline gives us, then, is not simply a metaphor for “data mining.” She gives us a narrative grammar for understanding what happens when an extractive system encounters a limit.
It looks for another frontier.
The hidden labour of the machine
That frontier is not always information in the conventional sense. AI systems also depend upon enormous amounts of human labour, much of it kept at the edge of public visibility. People classify images, compare responses, correct outputs, identify harmful content, and judge whether answers are useful, racist, abusive, sexual, violent, deceptive, or unsafe.
Some of that work has been outsourced into lower-wage labour markets, where workers have had to repeatedly encounter material involving sexual violence, racism, abuse, self-harm, and other disturbing content in order to make commercial AI systems safer for the people who eventually use them. Kenyan data worker Mophat Okinyi, describing his experience doing this kind of work for Sama on behalf of OpenAI, posed a brutally simple question: “Was my input worth what I received in return?”
The question punctures the fantasy that models somehow educate themselves. “The models just want to learn” may capture something about the technical dynamics of deep learning. But models do not go out into the world and satisfy their own curiosity. Institutions acquire compute, energy, data, land, labour, and expertise. Someone pays. Someone works. Someone is asked to surrender something. Someone else decides how much that contribution is worth.
The problem with the language of scale is that the larger the system becomes, the easier those relations are to disappear inside it.
The wrong apocalypse
This is what bothers me about the paperclip maximizer. The story imagines extraction as the symptom of a machine that has ceased to share human values. At some future moment, the AI becomes sufficiently powerful and insufficiently aligned. It stops recognizing the intrinsic value of human life. Everything around it becomes merely instrumental: bodies, buildings, ecosystems, planets, stars, raw material for the optimization function.
The machine is horrifying because it cannot see the world as anything other than resource.
But what, exactly, is alien about that?
Capitalism has spent centuries turning land, forests, minerals, bodies, attention, culture, and knowledge into inputs for accumulation. Colonialism repeatedly transformed inhabited territories into available land, relations into resources, and peoples into populations to be classified, administered, displaced, or absorbed. The machine in the paperclip story looks frighteningly strange only if we forget the histories that make its logic recognizable.
This does not make AI alignment unimportant. It changes the question. Instead of asking only how we prevent a future machine from treating the world instrumentally, we might also ask why the systems building that machine already depend upon forms of extraction we have learned to call normal.
The speculative future begins to look less like a rupture. It starts to look like an intensification.
The frontier moves inward
When I first gave the talk that became this essay, I was mostly thinking about scale moving outward: more books, more photographs, more websites, more conversations, more people, more of the world made available to computation.
I am less convinced now that outward is the only direction that matters. AI also scales inward.
Conversational systems can increasingly accumulate small traces of individual articulation across time: preferences, corrections, remembered details, recurring questions, stylistic habits, hesitations, associations, disclosures, and changes of mind. Their power does not only come from capturing more of the world. It can also come from increasing the resolution at which particular lives become computationally legible.
I have started thinking about these as two different directions of scale. Extensive scale is the familiar one: more data, more models, more users, more infrastructure. Intensive scale works through increasing resolution. It relates traces across time so that a particular person can become an increasingly detailed object of interpretation, prediction, personalization, and action.
The two movements are connected. Planetary infrastructures of computation and data make possible increasingly fine-grained forms of individualized address. Extensive scale moves outward across populations and archives. Intensive scale works across the accumulated traces of a particular life. The enormous infrastructure arrives at the interface as something that can feel remarkably small and intimate: a system remembering what you said months ago, noticing that you have returned to the same question, or recognizing the difference between what you said the first time and what you meant after correcting it.
The distinction matters because it changes what we imagine the next frontier of extraction to be. The frontier is not simply the rest of the internet or another untapped archive. It may increasingly be the fine-grained articulation of human life itself: the things we say while trying to understand ourselves, the questions we ask when we are frightened, the half-formed idea we ask a machine to help us develop, the conversation we rehearse before speaking to someone we love, the correction that reveals what we really meant, the pattern across hundreds of exchanges that neither we nor any individual interlocutor would have been able to hold together.
None of this means that a machine has discovered some authentic secret self buried inside us. That would reproduce precisely the fantasy I want to avoid. The issue is not whether AI can finally penetrate the human interior and extract its hidden truth. It is that more and more of what we experience as interior can be exteriorized through interaction, rendered persistent, related to other traces, and made computationally actionable.
That is a subtler form of enclosure.
And this is where I keep returning to The Marrow Thieves. The paperclip maximizer tells us to watch the machine. It asks us to imagine the terrible moment at which artificial intelligence ceases to share our values and begins converting the world into resources for its own purposes.
Dimaline directs our attention somewhere else. She asks us to watch the machinery of extraction: who needs the resource, who names the scarcity, who possesses the infrastructure capable of taking what supposedly cannot be replaced, and what relationships must disappear before something can become available for extraction.
And what happens when there is nowhere left for the frontier to move?
Dimaline’s answer is horrifying because the frontier does not disappear.
It moves inward.
Works Cited
Bostrom, Nick. Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014.
Coulthard, Glen Sean. Red Skin, White Masks: Rejecting the Colonial Politics of Recognition. University of Minnesota Press, 2014.
Dimaline, Cherie. The Marrow Thieves. Dancing Cat Books, 2017.
Hao, Karen, and Deepa Seetharaman. “Cleaning Up ChatGPT Takes Heavy Toll on Human Workers.” The Wall Street Journal, 24 July 2023.
Marx, Karl. Capital: A Critique of Political Economy. Vol. 1. Translated by Ben Fowkes, Penguin, 1976.
Patel, Dwarkesh. “Dario Amodei: $10 Billion Models, OpenAI, Scaling, & AGI in 2 Years.” Dwarkesh Podcast, 2024.
Villalobos, Pablo, et al. “Will We Run Out of Data? Limits of LLM Scaling Based on Human-Generated Data.” Epoch AI, 6 June 2024.
Younging, Gregory. “Gnaritas Nullius (No One’s Knowledge): The Public Domain and Colonization of Traditional Knowledge.” World Intellectual Property Organization.
