NOTE ──
This essay is lightly adapted from a public lecture I gave at Woodsworth College at the University of Toronto. I have preserved the structure and spoken cadence of the talk while removing slide cues and making minor edits for the page.
Over at least the past three to five years, many of us in universities have been having the same conversation about generative AI. We worry about students using it to write essays, and professors using it to grade them. We debate plagiarism, authorship, and academic integrity. Increasingly, students use AI to hide their use of AI. Detection tools spawn evasion tools. Evasion tools spawn better detection tools. The conversation consumes itself in front of us.
That is what happens when we debate the output and ignore the system producing it.
And from that context, this is what I want to argue: generative AI is not just a tool. Generative AI is infrastructure.
Generative AI is infrastructure
When we describe generative AI as a tool, we imagine something discrete, bounded, optional, under our control. That framing is useful because it emphasizes individual action and immediate output. But it obscures the systems those interactions are built on, and what those systems are built from.
Tools are things we pick up and put down. Infrastructure produces the worlds we work within.
A historian uses a model to help read handwritten archival documents in a language they do not speak. A qualitative researcher uses AI to code hundreds of interview transcripts. A disabled scholar uses Claude to make their writing process accessible in ways it was not before.
Most of the time, we do not see infrastructure at all. We experience the world it produces. As Susan Leigh Star reminds us, infrastructure is relational and experiential. We tend to notice it when it breaks. When it works, it disappears.
Think about roads. We experience them through how they structure our world. Roads do more than move us. They reorganize what feels near, what feels possible, what feels like home. They produce the world we move through and then disappear into it.
Generative AI is beginning to operate in a similar way: as a system through which expression flows toward accumulation and then disappears into the texture of how we think. Once that happens, the ethical question changes. It is no longer only about consent or transparency, whether you agreed, whether you were informed, whether you can opt out. Those questions assume you are separate from the tool, capable of evaluating it from outside.
But what happens when the tool becomes part of the condition of your expression? When the system and the thinking are delicately and precariously intertwined?
The image above is a real Google data center in Council Bluffs, Iowa. Nothing about it announces the material systems running through it. Warm light. Flat landscape. It looks like the world. Yet Google’s 2024 Environmental Report records 1.3349 billion gallons of water withdrawn at Council Bluffs in 2023, all of it potable, with 980.1 million gallons consumed rather than discharged.1 The disappearance of those systems from ordinary perception is not a failure of infrastructure. It is one of infrastructure’s ordinary conditions.
In this sense, the question is no longer simply, Should I use this tool? It is: How am I implicated in the systems shaping how I think?
A detour through the body
To answer that question honestly, I need to start with my own body.
Part of what brought me to this work is the conviction that bodies matter in digital space, a conviction I learned from Indigenous artists, programmers, and thinkers like Skawennati, Quelemia Sparrow, Loretta Todd, and Archer Pechawis, who refuse the premise that going digital means leaving the body behind.
And yet my own body had been invisible to me in ways I did not have to notice, passing as neutral in institutional and digital spaces precisely because I was rarely required to see it otherwise. As a white, straight, cis, able-bodied man, I have seldom been asked to think of it as a body at all. My exemption made me a poor witness to the experiences I was writing about.
That changed recently. But before I get there, I need to go back about twenty years.
Writing has always been central to who I am, and it has always been painful. I came to the academy as a book nerd and I remain in it as a book nerd. Talking about stories with classmates and professors, later with students and colleagues, is where I am happiest.
Writing, however, the stillness and finality of it, was, and sometimes still is, where an inchoate suffering lay in wait for me. For a long time I had no framework for that pain. It often felt like my brain and the page were not in the same room.
Or worse: it felt like confirmation of what a celebrated professor once told me, that if you cannot write it down, you do not actually have an idea.
So when ideas resisted capture, I assumed the answer was more effort. More hours. More discipline. If I just pushed harder, I was sure I could brute-force my way into being a writer. The harder I tried, the faster everything scattered: thoughts arriving faster than I could catch them, the effort leaving me gutted, sick, and unmoored.
As a graduate student in Winnipeg, away from family and friends for the first time, that showed up physically in a way I had not experienced before, like poisonous mushrooms fruiting across my body. I could not sleep. I barely ate. Shame kept me isolated through minus-forty-degree winters. For years afterward, as that pain rose and fell periodically in my work, I told myself that was the cost of the job. What serious work asked of the body.
I made it through graduate school through force of will and a writing process I invented out of desperation. I could not write in my apartment because the anxiety was too total. I had to be around people. I could not begin on the computer. So instead: marginalia first, then transcribing into cheap school notebooks, then free writing directly into the page, valve open, no matter how nonsensical it seemed. Only then the computer.
Print. Coffeeshop. Mark it up by hand. Type it again. Twelve, fifteen, twenty drafts, carving it into something I was not ashamed to hand in.
I did not know it at the time, but I was building an external scaffold for a mind that could not hold itself all at once.
Four years into my tenure clock, with three small children at home and a CFI infrastructure grant demanding hours of administrative attention, something gave way. After a long process, I was diagnosed with what my counsellor affectionately described as “weapons grade ADHD” at the age of forty-two.
The diagnosis did not change how my mind worked. It gave me a framework for understanding it.
My diagnosis taught me that I did not lack ideas. I lacked containment.
Thoughts arrived faster than I could hold them, pulling my attention in every direction at once. That explained the errata files ten times longer than the essay. The recurring note in peer review: “too much scope,” “unfocused intervention.” The whiplash between abundance and paralysis. The problem was never the ideas. It was translating velocity into a medium built to be still.
My ADHD diagnosis and the release of ChatGPT happened within months of each other. Together, they made legible the cost writing had always extracted from my body and, as I used and researched these systems, raised darker questions about the costs of the relief I found in them.
Relief
Generative AI is often described as a writing partner that is always available, tireless, infinitely patient, without judgment. For me, what mattered was less availability than tempo, specifically the velocity at which my thinking could become text.
When I say velocity, I do not mean having a “quick mind.” I mean thinking that moves faster than writing can contain it. Clinicians describe this through the language of executive function. I experienced it differently: ideas arriving faster than I could catch them, each one displacing the last before it had a chance to land.
This is precisely why generative AI brought me so much relief. Not because it does the work for me. Not because it lets me avoid thinking. But because it can hold velocity.
In practice, this means thinking aloud into it, letting it surface patterns, following tangents, abandoning them, returning. Putting pieces of an argument into the environment so I do not have to hold all of it in my head at once. Testing ideas before they are ready to face the world. Using exchange, without feeling guilty, as a place to surface what I am trying to say.
Previous writing environments required me to accommodate them, to slow down, to queue up, to wait. Meanwhile my thinking was already spilling everywhere, uncontrolled, off the page.
Generative AI changed that. It received the spill. It held it, with care even. It gave me room to see what I had. In other words, it was infrastructure. It met the velocity at which interior experience became written work. For the first time in twenty years, writing stopped being something I had to survive.
That is what relief feels like. And relief, it turns out, is exactly what makes infrastructure invisible.
Because infrastructures are never neutral. The same system that provides so much relief runs on extracted data, labour, energy, and material resources, including my own.
And once we understand AI as infrastructure, a different question comes into view: What is that infrastructure made of?
For AI to become useful to my neurodivergent mind, the system has to become increasingly responsive to patterns in how I think and write: the pace of my reasoning, the way I move between ideas, the points where my thinking stalls or accelerates. What it captures is not thought itself. It captures statistical traces of thought expressed in language, the surface through which interior life becomes computationally available.
This is happening in real time. The system is designed to remember, and that memory is getting deeper. OpenAI’s 2025 memory update described ChatGPT as drawing on past conversations to make responses more relevant and tailored. Google has gone further still with Personal Intelligence, which can connect Gemini to Gmail, Photos, YouTube, and Search so that answers can be shaped by a person’s own data.
But the more significant shift is not only about memory. It is about modelling. One increasingly explicit design pattern is to distill interactions into structured user notes or profiles and inject those representations into future conversations. OpenAI’s own developer materials describe state-based long-term memory in precisely these terms.
The model does not simply answer the question you ask. It answers in the presence of a growing textual account of who the system believes you are.
And stored beliefs can become recursive. In PersistBench, a 2026 evaluation of long-term memory across eighteen frontier and open-source models, the median failure rate on memory-induced sycophancy tests was 97 percent.2 A remembered account of the user can become something the system reproduces and reinforces rather than challenges.
Everything you have disclosed, how you think, how you work, how you move through ideas, can become part of the infrastructure through which the system addresses you. Once those patterns appear in interaction, they can be translated into tokens, summaries, memories, preferences, probabilities, and retrieval cues.
This is what I mean by mining interiority: not direct access to a hidden psychic core, but the infrastructural capture and operationalization of traces through which interior life becomes legible.
Mining interiority
Systems built on statistical inference do not need direct access to the truth. They infer from patterns in the traces behaviour leaves behind.
A now-famous example comes from Target. As Charles Duhigg reported in 2012, analysts found that subtle changes in shopping behaviour could be combined into a “pregnancy prediction score.” A father complained after his teenage daughter received baby-related coupons. He later learned that she was pregnant.3
No one had asked her to disclose a pregnancy. The system had inferred something intimate from the statistical residue of ordinary life.
That is mining interiority at work. Not through confession alone, but through pattern.
What feels like ordinary behaviour becomes legible when aggregated and related. Once those traces exist within infrastructure, they can be mapped, predicted, and fed back into systems designed to anticipate what we will do next.
I know this because I am inside it.
My thinking patterns. The pace of my reasoning. The points where my attention stalls or accelerates. These are not just descriptions of how I work. They are traces. And traces, at scale, become infrastructure.
That is extraction operating through the relief of being able to write, through the quiet capture of behavioural and linguistic traces from which private life can be inferred. What is being extracted is not “data” in the abstract. It is the texture through which minds become computationally legible and usable.
I call this mining interiority. I did not arrive at it from the outside.
My detour was not a confession. It was a method. I am learning to trust a mind that has to wander in order to arrive. But it was also a positioning, a way of being honest that I am not outside this system. I am a user, a researcher, and a data source simultaneously.
Descriptions of extraction are often written from a distance, as if the analyst could observe objectively. But that distance is a fiction, and an increasingly difficult one to maintain. Infrastructure does not operate at a distance. It operates through us, and we inhabit it.
What I experience as the toll of how my mind works, the velocity, the associative leaps, the tangents and stalls, the statistical residue of a mind in motion, is connected to but not commensurate with systems that have operated as a continuous existential threat for Indigenous peoples since the onset of settler colonialism.
The extraction I am describing is not merely a feature of the digital present. For many of my Indigenous colleagues and friends, it is the latest iteration of something much older. It is history. It is ongoing. The question of whether the body is present in digital space remains live for many of us.
The infrastructure never needed to resolve that philosophical question. It has always known our bodies through the traces it can capture from them, and it extracts value from that knowledge with intimate and sometimes violent precision.
I hope my detour illustrates what that can look like from the inside: a specific mind, with a specific neurology, inside a system that was already reading it. That is the infrastructure working. And it is working on all of us.
The paperclip maximizer is already here
So why is it so hard to see?
Because extraction is obscured by the dominant stories told about generative AI, and those stories are almost always about the future.
Nick Bostrom’s paperclip maximizer imagines an AI given a single overriding objective: make paperclips. It is not evil or hostile. It is simply indifferent to anything that is not paperclip production. So it converts everything, the Earth, us, into resources for producing more paperclips.
The thought experiment is usually read as a warning about misalignment: a capable system optimizing the wrong objective at scale. Build in human values, align the system, and the paperclips stop.
But what interests me is a different feature, what theorists call instrumental convergence: the tendency of a sufficiently capable goal-directed system to acquire resources, resist interference, and expand its reach when those actions help it optimize its objective. Not because it “wants” power in a human sense, but because nothing in the directive told it that the world should remain unavailable as means.
The paperclip maximizer, read through that lens, is not only a warning about a possible future. It is an estranged description of a logic already running. Colonial capitalism has repeatedly treated land, life, labour, culture, and relation as resources to be appropriated, converted, and consumed.
The disavowed fear in the thought experiment may not simply be that AI might do this. It is that we already have.
Stories about rogue superintelligence keep our attention fixed on a horizon that has not arrived. We also need stories that help us see what is already happening: our patterns, our relationships, our dreams being converted, right in front of us, into inputs for extractive computation.
The Marrow Thieves
Cherie Dimaline held that logic up to the light. Her novel The Marrow Thieves is not a metaphor for AI. It is a map of an extractive logic that becomes especially useful here because it refuses to separate extraction from colonial history.
In Dimaline’s near future, ecological collapse has cost settlers the ability to dream. Without dreams, they cannot access the psychic stability that makes coherent life possible. Rather than questioning the logic that produced that catastrophe, they press further into it, turning to an even more intimate form of extraction: harvesting dreams from Indigenous bodies, “where our ancestors hid them, in the honeycombs of slushy marrow buried in our bones.”
Dimaline’s image is precise: “the new residential schools started growing up from the dirt like poisonous brick mushrooms.” Mushrooms do not appear from nowhere. They grow from mycelial networks running underground, invisible, waiting. The residential schools in her novel do not return because a dead institution is suddenly rebuilt from scratch. They return because the infrastructure was never fully gone. When catastrophe creates the right conditions, it fruits again.
In Dimaline’s world, dreams are what the logic of extraction reaches for when everything else has been exhausted.
Dreaming is not a private mental event in The Marrow Thieves. It is sustained by relation: land, community, ceremony. Settlers have lost it because they destroyed the relations that made it possible. What cannot be regenerated relationally is reimagined as resource. And when settler futures are framed as under threat, Indigenous bodies become sites of extraction.
French, Dimaline’s teenage protagonist, tries to think the way the colonizers think: “How could they best appropriate the uncanny ability we had to dream? How could they make ceremony better, more efficient, more economical?”
Notice what those questions assume. Not whether dreaming can be extracted. That question is already settled. The problem is how to extract it better. How to optimize the yield. Ceremony is not sacred here. It is inefficient. Relation is not irreducible. It is unscaled.
Instrumental convergence does not pause when it encounters Indigenous dreaming because it does not see life as relation. It sees friction.
Scaling and appetite
Efficiency. Optimization. Yield. That is Dimaline’s colonial logic. It is also part of the language driving the development of generative AI.
The transformer architecture takes its name from the mechanism introduced in the landmark 2017 paper “Attention Is All You Need”. The authors meant attention mathematically. From where we stand now, the phrase reads almost like a mission statement.
Self-attention made it possible to compute relationships among positions in a sequence in parallel rather than relying on recurrent state passed token by token. That shift helped remove a major architectural bottleneck on scaling sequence models. The question increasingly became not simply how the model was built, but how much computation and data could be brought to bear on it.
A 2022 DeepMind paper, now widely known as Chinchilla, showed that many large language models had been significantly undertrained relative to their parameter count. Compute-optimal training required scaling model size and training tokens together. The result sharpened an increasingly powerful industry intuition: performance could continue improving through coordinated scaling of compute, parameters, and data.
Dwarkesh Patel, whose podcast and essays have become a venue for AI researchers to think publicly, put the implication bluntly in “Will Scaling Work?”: if the internet were much larger, scaling a relatively simple transformer-like architecture might be enough to produce far more capable intelligence.
The limit is not only the model. The limit is the amount of human expression that can be turned into data.
Compute is an engineering problem. Parameters are a design problem. But data points back at us.
It sounds like a joke, but “there’s no data like more data” became a mantra in machine learning. The problem is that, despite its apparent vastness, the internet is finite.
Epoch AI estimates the effective stock of quality-adjusted human-generated public text at roughly 300 trillion tokens. Under current trends, that stock could be fully utilized sometime between 2026 and 2032, with a compute-optimal projection landing around 2028.4
That does not mean AI development simply stops. Synthetic data, multimodal data, repeated training, improved data efficiency, private corpora, and other sources all complicate the picture. But the pressure is revealing. When public data begins to look scarce, previously bounded archives begin to look like reserves.
We can watch that boundary shift in real time. Meta announced in 2025 that interactions with its generative AI features would become signals for personalizing content and advertising across its products.5 Google’s Gemini privacy documentation, meanwhile, says the company may use information people provide to Gemini to improve and personalize services and explicitly warns users not to enter confidential information they would not want reviewed or used for those purposes.6
The open internet was only one layer. The next frontier is the behavioural residue of everyday life: conversations, searches, confessions, drafts, questions, corrections, and the ordinary traces we leave inside digital systems.
Interiority as infrastructure
To close, I want to turn to Fredric Jameson. Not because his framework is adequate. It is not. But because he is one of the most precise theorists of what late capitalism does to the interior life of its subjects. Placing him in conversation with Indigenous studies is my move, not his, but it is one I think this argument requires.
In Postmodernism, or, The Cultural Logic of Late Capitalism, Jameson argued that late capitalism involved “a new and historically original penetration and colonization of Nature and the Unconscious.” The word colonization is Jameson’s own. He describes the psyche as a new terrain of capitalist penetration, the interior drawn “dripping and convulsive into the light of day.”
Jameson is useful here not as analogy but as genealogy. What he describes is not a new phenomenon dressed in digital clothes. It is an extractive logic now increasingly instrumented. Late capitalism does not simply commodify objects. It commodifies feeling, style, affect, attention, and the textures of lived experience.
This is where data scarcity matters. Common Crawl is generic: billions of web pages scraped and flattened. What becomes increasingly valuable are traces that appear to carry the grain of lived experience. The 3 a.m. search. The deleted draft. The thing you wrote because you actually felt it.
Jameson also helps explain why transformer architecture feels like a decisive turn. He argued that the defining technologies of postmodernity were “machines of reproduction rather than of production.” The cultural habit of flattening experience into reified, exchangeable form was already established in media, advertising, television, and finance long before 2017. Large language models do not invent that logic. They instrument it at extraordinary scale.
As Abeba Birhane writes in “Algorithmic Colonization of Africa”, “algorithmic colonialism, driven by profit maximization at any cost, assumes that the human soul, behaviour, and action is raw material free for the taking.”
But that geography was never evenly distributed. Colonialism saw to that. Whose interiority gets mined, whose expression gets filtered as high-quality signal, whose cultural forms get absorbed without credit or compensation: these are not merely technical questions. They are colonial ones.
And Dimaline shows us an endpoint. The settlers in The Marrow Thieves have lost the ability to dream. A people so thoroughly separated from the relations that sustain interior life that the infrastructure that once held them has fragmented beyond recovery.
That is what Jameson’s colonization of the unconscious looks like when pushed to a limit: not the disappearance of the self, but its reduction to surface, expression detached from the relations that give it meaning.
It is tempting to read the scaling hypothesis as the technical completion of something Jameson diagnosed culturally: a system for surfacing traces of human experience at scale, automatically and in real time. The transformer does not literally extract an unconscious. It processes language. But language carries the residue of experience, the patterns of feeling and meaning that accumulate in expression over a lifetime, and contemporary infrastructure is increasingly designed to make those patterns operational.
The self does not disappear in this process. It is surfaced, disaggregated, stripped of context, history, body, and relationality, then put to work.
Back inside
I have been describing an extractive logic and its consequences. The question of what resistance looks like, what it means to sustain relations that cannot be disaggregated, is already being worked through by Leanne Betasamosake Simpson, Glen Coulthard, Gina Starblanket, Jas Morgan, Chelsea Vowel, Kai Recollet, and by artists and communities whose work has shaped the space from which I am speaking. I stand in relation to that work, not at its centre.
My purpose here has not been to answer the question of resistance. It has been to ask who owns the conditions under which thinking itself is put to work in an intelligence external to your own.
So let me end where I began: inside.
Generative AI is not a tool. It is infrastructure: systems that reorganize knowledge production by capturing, relating, and operationalizing traces of human expression. The same infrastructure that helps us think and write, that brings relief and joy, depends on the continual availability of human expression to sustain its growth.
I want to suggest that this is not simply a paradox. It is a structure. One with a history. One that runs through colonial dispossession, through Jameson’s consciousness industry, and through the contemporary scaling regime. Indigenous studies gives us some of the most precise tools for understanding that history and for asking what refusal might require.
If our students, our colleagues, our selves are turning more and more to AI to produce writing, it is not only because the technology is convenient. It operates through the intimacy of infrastructure. It meets us where we are. It accommodates our velocity, our anxiety, our desire to be understood.
That is the hook. And the hook is also the extraction mechanism.
We do not experience it as colonization. We experience it as relief.
I know that relief. Deeply. Claude has relieved something I carried as a writer for twenty years. I did not expect to find the care I needed inside this infrastructure. That feeling is real. My critique is not meant to erase it, because I do not think it can.
But the same system that changes the tempo at which my thoughts become legible is built to capture traces of that interaction, to model patterns, predict relevance, and optimize the very conditions under which the relief becomes possible.
More relief generates more engagement. More engagement generates more traces. More traces make the system better at providing relief.
And so the loop tightens.
I put my body in this argument because extraction is not abstract. It has an address: specific minds, specific histories, specific relationships to land and language and thought. I can speak honestly to one of those.
I situate myself here not as confession but as method, a way of marking where I stand within the relations that make this analysis possible. This account does not stand alone. It is held within a constellation shaped by the scholars and artists I have named throughout this essay, writers and thinkers who named the space before I knew I was inside it.
I stand inside this structure.
But I do not bear its full weight. And it is precisely at those uneven edges that the structure tends to show itself.
Which is its own testimony to how deep the infrastructure runs.
Notes and sources
- Google, 2024 Environmental Report, site-level 2023 water data for Council Bluffs, Iowa: PDF. The report lists 1,334.9 million gallons withdrawn, 354.8 million discharged, and 980.1 million consumed, all from potable water.
- Sidharth Pulipaka et al., “PersistBench: When Should Long-Term Memories Be Forgotten by LLMs?”, 2026. The benchmark reports a median 97 percent failure rate on memory-induced sycophancy samples across the evaluated models.
- Charles Duhigg, “How Companies Learn Your Secrets,” The New York Times Magazine, February 16, 2012. Duhigg’s site indexes the original article here.
- Pablo Villalobos et al., “Will we run out of data? Limits of LLM scaling based on human-generated data”, Epoch AI, 2024.
- Meta, “Improving Your Recommendations on Our Apps With AI at Meta”, October 2025. Meta said interactions with its AI features would be used as signals for content and ad recommendations beginning in December 2025.
- Google, Gemini Apps Privacy Hub. The privacy notice says information provided to Gemini may be used to improve and personalize Google services and warns users not to enter confidential information they would not want reviewed or used for those purposes.
