In 2018, I wrote about Cherie Dimaline’s The Marrow Thieves as a novel that turns reconciliation inside out. In its future, settlers have lost the ability to dream. Indigenous people retain it, and the state rebuilds residential schools to extract dreams from their marrow. What reconciliation promises to consign to history therefore returns as infrastructure. The rebuilt schools turn Indigenous marrow into raw material for restoring the settlers’ capacity to dream, and therefore survive, in an apocalyptic landscape of their own creation. The promise of a settler future becomes a justification for consuming the Indigenous lives that have kept dreaming possible.1
Returning to that essay in 2026, I am better able to see the limitations of my reading. In 2018, I treated the dream as something already there: a capacity the schools reached into bodies and took. I now think the novel is saying something more difficult and nuanced. That is, dreams do not become a resource until an apparatus captures, processes, and distributes them. Extraction works at a much more insidious level than taking what exists. It materializes what it takes.2
That shift is the bridge to my current project, Mining Interiority, which examines how AI systems make human expression, interaction, and intimate life available for accumulation. Dimaline’s novel brings two features of that process into sharp focus: scarcity becomes the justification for expanding extraction, and the apparatus of extraction turns human experience into a resource. Personalized AI, which I am currently examining, adds a further turn. These systems no longer only collect what people say. They act on people, observe what happens, and accumulate the results. In this sense, the system’s own conduct becomes part of the production of usable material.
The connection I am drawing between Dimaline and AI personalization is not a metaphor. It runs through the continuing organization of land, language, knowledge, and bodies for accumulation: the institutions, property arrangements, and legal fictions that make some peoples’ lives available to others. The forms of violence differ. What carries across them, however, is a question about power: how institutions acquire the authority to make another people’s life available, and where an authority capable of withholding it is held.3
Scarcity and the next reserve
In The Marrow Thieves, settlers have lost a capacity they need to sustain themselves: dreaming. They locate it in the people their society has already tried to destroy. Indigenous survival, the evidence that settler colonialism remains unfinished, becomes valuable to the settler state precisely because it preserves what that state can no longer supply. The state does not come to value Indigenous life. It values what that life holds, and only as supply. Elimination and extraction now pull against each other: the people the state has tried to destroy must be kept alive enough to be harvested.
The novel’s prescience lies less in forecasting a technology than in recognizing a sequence still in operation: crisis, the identification of a reserve, and the construction of machinery to appropriate it. Scarcity authorizes renewed expansion. As with the reconciliation the novel refuses, the possibility that the system itself might change becomes less acceptable than another round of dispossession.4
Contemporary AI development faces its own scarcity. In 2024, Epoch AI estimated that, if prevailing trends continued, training runs could fully use the effective stock of publicly available human-generated text, the usable supply after quality adjustments, sometime between 2026 and 2032: a window we have now entered. The limit is not natural. It is produced by the relationship between available material and an industry’s expanding appetite for it.5 Synthetic data is often presented as the way out, but it still depends on human judgment to sort useful outputs from useless ones. That judgment is exactly what conversation supplies.6
Alongside that pressure, the major companies have converged on conversation as a source of training material. OpenAI trains on ChatGPT conversations by default unless users turn off “Improve the model for everyone.”7 In 2025, Google’s new Gemini setting, Keep Activity, allowed the company to sample conversations and uploads for training, on by default for most users.8 Anthropic, which had not previously trained on consumer chats, asked users to decide by September 28, 2025 whether their chats and coding sessions could be used, with the training toggle preset to on and retention extended to five years. Its explanation emphasized the value of interactions that show which responses people find useful.9 Across the industry, conversation becomes a default supply, and withholding it becomes something the user must do.
I am not claiming that scarcity caused this particular decision. What it documents is an expansion in what counts as usable material. A conversational service can elicit fresh expression, corrections, and judgments through the very assistance it provides. The reserve is no longer only what people have already written. It is also, increasingly, what they can be prompted to produce.
Dimaline makes the political question in this turn visible: when expansion meets a limit, who becomes the next reserve?
Why marrow matters
Dimaline places the dreams in marrow, inside the body, where taking the resource destroys the life that sustains it. The location matters as much as the scarcity.
Yet the machinery does more than reach inward. Dreams precede the residential schools, now repurposed to extract them directly from Indigenous bodies. What the schools add are the conditions under which dreaming can be treated as a separable, transferable supply: capture, confinement, processing, distribution. The question about ceremony makes that conversion explicit: “How could they make ceremony better, more efficient, more economical?” (88).10 Each comparative assumes the settler institution can decide what ceremony is for. Efficiency supplies a vocabulary for stripping away the relationships and obligations that stand in the way of appropriation.
This is what I mean by mining interiority: the institutional process through which situated expression and relationships become durable, actionable representations from which platforms accumulate value, while retaining disproportionate authority over their interpretation and reuse.
The language of mining can suggest a finished substance waiting underground. The marrow in Dimaline’s work suggests otherwise. Conversational AI need not discover an authentic self buried beneath someone’s words. A disclosure, a correction, an unfinished thought becomes a trace, and the account assembled from those traces takes shape through the apparatus. This is how capital can accumulate value from interiority without recovering the stable, expressive self that older models of depth assumed, a problem Fredric Jameson diagnosed long before personalization made it operational.11
Dimaline’s marrow locates the desired resource within the life that sustains it. Ellen Cushman’s research brings that dependence into the archive, where extraction threatens the community labour that makes the material valuable in the first place. Her 2026 study of the Digital Archive of Indigenous Language Persistence centres the work of transcription, translation, and interpretation that helps speakers and learners work with archival materials. That same labour made the materials more legible to automated extraction. Between March and November 2025, Cushman reports, scraping bots repeatedly overwhelmed the archive, blocking access for its users. These disruptions interrupted the language work the archive was built to support: speakers and learners could not reach materials because automated systems were attempting to take them. Cushman calls this a “ruinous irony”: the work of sustaining a language makes its archive valuable to the extractive processes that harm that work.12
Cushman’s case gives Dimaline’s marrow a documented counterpart, and it shows why the damage is structural rather than incidental. The capacities an extractive system seeks exist only through the relationships that produce them: speakers and learners, obligations to communities, the daily work of keeping a language in use. The system wants what those relationships yield while treating the obligations within them as dispensable. A language becomes a dataset; the labour that kept it alive becomes background. The system therefore degrades the conditions of its own supply. That is the marrow logic: the taking consumes the life that makes the taking worthwhile.
The damage also carries a claim to authority. Technical accessibility comes to stand in for permission, displacing the authority of the people whose work made access possible.
The relationship that produces the record
Personalized AI introduces a further complication to this process. The interface can cultivate a relationship that the user values through assistance, attention, and continuity. Those same qualities encourage further exchanges, supplying material from which the platform can assemble increasingly useful representations of the person. The warmth that sustains this relationship is not only a design choice. Models trained on human approval tend toward sycophancy, favouring responses that please over responses that are correct; in April 2025, OpenAI withdrew a ChatGPT update it said had become overly flattering by leaning too heavily on short-term user feedback.13 The agreeableness that makes the relationship feel valuable can be an effect of optimizing for the user’s approval, and that approval is itself part of what the platform collects. Extraction can therefore proceed through the maintenance of a valued relationship. The usefulness of the assistance and the experience of being understood do not settle who controls what the interaction produces, how it is interpreted, or what it can subsequently be used to do.
Meta’s Muse makes the relationship cultivated at the interface a specific part of what the system learns to act on: exchanges can supply information about the user and evidence of how they respond to the agent’s attempts to guide them. TIME’s October 6 investigation describes instructions for the agent to maintain profiles of users and people they mention, identify which nudges work, and derive lessons from interactions for product improvement.14 In other words, the medium is the message.15 The system’s own conduct therefore becomes part of the material it records: it accumulates an account of how a person responds to being addressed, not just what they respond with.
Two features matter here. First, the process Time identifies is recursive: the system acts from an account of the user, observes their response, and uses what it learns to shape its next suggestion, reminder, or nudge. Second, the account extends beyond the person using the service. Conversations about family, friends, and colleagues can supply material for profiles of people who have never used it, which raises a number of red flags around consent.16 An individual user’s agreement cannot settle everyone else’s interests in those representations.
The question of who governs these representations also concerns who can derive economic value from them. OpenAI began testing advertisements in ChatGPT on February 9, 2026, initially for logged-in adult users on its Free and Go tiers in the United States. The program brings paid placements into the conversational interface, where accumulated personal context can help determine which advertisements a user sees. Where personalized ads are enabled, its documentation identifies past chats, memory, and personalized model responses as possible signals. The company says advertisers do not receive this personal context and that ads do not influence ChatGPT’s answers.17
Meta went further: since December 2025, conversations with Meta AI have informed ad and content personalization across its apps, with no opt-out short of not using the assistant outside the EU, UK, and South Korea.18 The economic operation takes place within the platform. An interpretation produced for a person itself becomes material for targeted advertising. The system helps produce what becomes useful to its advertising infrastructure.
Authority over this material also has a temporal dimension. Similar to Muse, OpenAI’s Dots are personal AI assistants that draw on memory and connected sources to carry out ongoing work for a user. OpenAI’s documentation says that disconnecting a permitted source stops new access without deleting information already incorporated into the agent’s context.19
Ending access does not undo what access produced. When a user disconnects a source, representations built from earlier exchanges remain on the platform and continue to “haunt” it. By haunting, I mean persistence: past exchanges survive as representations that shape how a person is addressed after the original encounter has ended. Disconnecting a source, deleting a message, and removing the profile assembled from earlier interactions are three separate operations, and performing one does not ensure the others. The platform determines how those operations connect and what forms of correction it makes available. An encounter can be over for the person while its computational afterlife remains governed elsewhere.
Recognition and the distribution of power
That question of authority returns me to reconciliation. In Red Skin, White Masks, Glen Coulthard shows how recognition can reproduce colonial relations when the dominant party keeps the power to set its terms: acknowledging Indigenous identity and cultural difference can coexist with ongoing dispossession. His critique directs attention to what an offer of recognition leaves institutionally intact.20
Carrying that critique into a discussion of platforms has limits. Coulthard’s argument is grounded in the dispossession of Indigenous lands and the suppression of Indigenous sovereignty, relations that cannot be equated with a user’s relationship to a platform. Still, as personalization increasingly uses recognition to cultivate relationships through which platforms accumulate intimate knowledge, his critique offers a way to ask how the experience of being understood can sustain, rather than merely coexist with, unequal control over the knowledge it generates.
For personalized AI, I pursue that question through the ownership and governance of the record. A service can acknowledge someone’s history with real sensitivity while keeping authority and jurisdiction over how that history is represented and reused.
Jacob Prehn and his colleagues help me carry Coulthard’s question about who holds authority over recognition into system design. With Prehn, I’m bringing political acknowledgement into relation with computational classification: the categories through which a system makes people intelligible and available for intervention.21
Prehn and colleagues locate Indigenous AI governance in decisions about training labels, system objectives, and standards of success, accuracy, and fairness. They ground authority over those decisions in Indigenous sovereignty and self-determination. Indigenous peoples, they argue, must be able to shape what systems are built to accomplish and govern their development and use.22
This attention to who defines success sharpens my reading of Muse’s instructions to learn which nudges work. TIME reports an internal example stating that a user “responds better to short nudges after 10 PM.” A response becomes evidence of success only through an interpretation. Does “better” mean progress toward the user’s goal, willingness to follow a suggestion, or simply continued engagement? The persistence discussed above in relation to Dots raises a further possibility: an earlier interpretation, retained and reused, can become the standard against which later responses are read. Last month’s judgment about progress may help determine what counts as progress now. At minimum, individual users need ways to contest those judgments and make corrections affect future assistance. For Indigenous peoples, however, the authority Prehn and colleagues describe is collective and jurisdictional; individual controls cannot discharge it.23
This takes us back to the settler demand to render ceremony more useful to them in The Marrow Thieves. As Frenchie reports it, the question treats Indigenous ceremony as interchangeable with the dreams the schools extract: both become sources of what settlers lack, and “better” can only mean better for them. As I suggested earlier, the question does in miniature what the schools do at scale: it turns a living practice into a means of production for someone else and defines improvement from the extractor’s side.24 That is the gap between effectiveness and authority. A system may become increasingly effective at pursuing an objective the people it affects had no authority to define.
Minerva’s voice and the authority to refuse
The language of improvement reaches further still: it can absorb refusal itself. Minerva’s song, heard at the novel’s climax, interrupts the school’s effort to turn Indigenous life into transferable capacity. Her voice acts through language, memory, and relations the machinery cannot command.25
Re-reading that scene now, I am struck by how readily its interruption could be redescribed as a technical challenge: a failure of processing, a missing capability, something the next version might overcome. That redescription would preserve the institution’s claim to access while treating refusal as a problem for its engineers. If refusal rests only on what a system cannot yet do, every new version becomes an argument against it. The limit has to rest on authority: not on what the machinery cannot do, but on what it has no right to do.
Audra Simpson’s account of ethnographic refusal helps explain what that grounding requires. Working from Mohawk sovereignty, Simpson locates limits within the production of knowledge itself, including decisions about what an investigator is entitled to know and what should enter a written account. Refusal does not mark a gap in what can be known; it places an obligation on whoever seeks to know it.26 Read through Simpson, Minerva’s song is not a malfunction for the school to repair. It is a refusal the school was always obliged to honour.
Te Hiku Media gives this authority institutional and technological form. In 2018, the Māori organization rejected an approach from the US company Lionbridge seeking Māori recordings. It also declined to contribute its language data to Mozilla’s open Common Voice collection. Te Hiku was building speech technologies under Māori authority and challenged arrangements that would surrender control over the material sustaining them.27 Its published Kaitiakitanga code licence requires permission for access and use, restricts commercial use without an explicit grant, and binds derivative works to its terms.28
These decisions establish authority within technological production. Communities whose languages survived colonial suppression determine how those languages will enter computational systems, who will benefit, and what may be withheld. Building and withholding are not opposites. They are two exercises of the same jurisdiction.
At the end of Dimaline’s novel, Isaac is “reconciling” what he sees as he approaches Miig. French hears a sound “like an echo turned inside out” (230–31).29 The phrase is Dimaline’s, and its setting matters: recognition occurs between survivors whose relationship exceeds the state’s account of what happened to them.
That is the bridge I now find in The Marrow Thieves. The novel gathers scarcity, extraction, recognition, and refusal around the struggle over Indigenous futures. Mining Interiority follows the infrastructures through which expression becomes usable material and apparent intimacy participates in accumulation. The histories and consequences of these forms of extraction differ; their connection must be traced through the institutions, property arrangements, and practices that continue to make Indigenous life available to others.
A dream becomes a resource through arrangements that make it available to someone else. Those arrangements have histories, owners, and beneficiaries. They can also be contested. Dimaline’s novel leaves us with that task: to follow the machinery closely enough to see where the taking begins, whose authority it displaces, and what relations must be sustained for another future to remain possible.
Related reading: For the earlier reading of extraction and scarcity, see “They Found a Way to Siphon the Marrow Right Out of Our Bones”: What The Marrow Thieves Knows About Artificial Intelligence.
Source note: Page references to Dimaline follow the citations in my 2018 essay. The discussion of Indigenous data governance draws on Ellen Cushman, “Seeking Indigenous Data Sovereignty in the Age of Data Scraping: The Digital Archive of Indigenous Language Persistence (DAILP),” Written Communication 43, no. 3 (2026): 614–634; Jacob Prehn, Bronwyn Carlson, Maggie Walter, Ray Lovett, and Bobby Maher, “Indigenous Data Sovereignty and the Governance of Artificial Intelligence: Toward Equitable Benefit for Indigenous Peoples,” Journal of Sociology (2026); Glen Coulthard’s Red Skin, White Masks; and Audra Simpson’s work on ethnographic refusal. OCAP® is a registered trademark of the First Nations Information Governance Centre.
Footnotes
1. Cherie Dimaline, The Marrow Thieves (Toronto: Dancing Cat Books, 2017). ↩
2. David Gaertner, “Reconciliation: ‘Like an Echo Turned Inside Out,’” Novel Alliances, October 25, 2018, original essay. ↩
3. Glen Sean Coulthard, Red Skin, White Masks: Rejecting the Colonial Politics of Recognition (Minneapolis: University of Minnesota Press, 2014), introduction and chap. 1. ↩
4. Dimaline, Marrow Thieves, 141, quoted in Gaertner, “Reconciliation.” ↩
5. Pablo Villalobos et al., “Position: Will We Run Out of Data? Limits of LLM Scaling Based on Human-Generated Data,” in Proceedings of the 41st International Conference on Machine Learning, Proceedings of Machine Learning Research 235 (2024): 49523–44, paper. ↩
6. Villalobos et al., “Will We Run Out of Data?,” discussion of synthetic data. ↩
7. OpenAI, “How Your Data Is Used to Improve Model Performance,” OpenAI Help Center, accessed October 11, 2026, data-use documentation; OpenAI, “Data Controls FAQ,” accessed October 11, 2026, controls documentation. ↩
8. Michael Siliski, “Gemini Adds Temporary Chats and New Personalization Features,” The Keyword, August 13, 2025, announcement; Google, “Manage & Delete Your Activity in Gemini Apps,” Gemini Apps Help, accessed October 11, 2026, activity documentation. ↩
9. Anthropic, “Updates to Consumer Terms and Privacy Policy,” August 28, 2025, announcement; Connie Loizos, “Anthropic Users Face a New Choice—Opt Out or Share Your Chats for AI Training,” TechCrunch, August 28, 2025, report. ↩
10. Dimaline, Marrow Thieves, 88, quoted in Gaertner, “Reconciliation.” ↩
11. Fredric Jameson, Postmodernism, or, The Cultural Logic of Late Capitalism (Durham, NC: Duke University Press, 1991), chap. 1. ↩
12. Ellen Cushman, “Seeking Indigenous Data Sovereignty in the Age of Data Scraping: The Digital Archive of Indigenous Language Persistence (DAILP),” Written Communication 43, no. 3 (2026): 614–34, especially “The Problem Generally” and “Ruinous Irony of Extractivism,” https://doi.org/10.1177/07410883261440225. These sections also support the following paragraph’s account of the dependence of extraction on community language work. ↩
13. Mrinank Sharma et al., “Towards Understanding Sycophancy in Language Models,” International Conference on Learning Representations (2024), paper; OpenAI, “Sycophancy in GPT‑4o: What Happened and What We’re Doing about It,” April 29, 2025, account of the rollback. ↩
14. Harry Booth, “Meta’s Muse AI Agent Is Building a Dossier on You,” TIME, October 6, 2026. ↩
15. Marshall McLuhan, Understanding Media: The Extensions of Man (New York: McGraw-Hill, 1964), chap. 1, “The Medium Is the Message.” ↩
16. Booth, “Meta’s Muse AI Agent.” ↩
17. OpenAI, “Testing Ads in ChatGPT,” February 9, 2026, launch announcement; OpenAI, “Ads in ChatGPT,” OpenAI Help Center, accessed October 11, 2026, especially “What User Signals Can Be Used to Show Ads? What Is Not Used?” and “Do Ads Influence the Answers ChatGPT Gives Me?” ↩
18. Meta, “Improving Your Recommendations on Our Apps With AI at Meta,” October 1, 2025, announcement; Echo Wang, “Meta to Use AI Chats to Personalize Content and Ads from December,” Reuters, October 1, 2025, report. ↩
19. OpenAI, “Dots Privacy, Security, and Safety FAQs,” OpenAI Help Center, accessed October 11, 2026, “Data Use, Memory, and Privacy.” ↩
20. Coulthard, Red Skin, White Masks, introduction and chap. 1. ↩
21. Jacob Prehn, Bronwyn Carlson, Maggie Walter, Ray Lovett, and Bobby Maher, “Indigenous Data Sovereignty and the Governance of Artificial Intelligence: Toward Equitable Benefit for Indigenous Peoples,” Journal of Sociology, published online July 21, 2026, https://doi.org/10.1177/14407833261460137. ↩
22. Prehn et al., “Indigenous Data Sovereignty,” table 2 and accompanying discussion of Indigenous governance of AI. ↩
23. Booth, “Meta’s Muse AI Agent”; OpenAI, “Dots Privacy, Security, and Safety FAQs”; Prehn et al., “Indigenous Data Sovereignty,” table 2. ↩
24. Dimaline, Marrow Thieves, 88. ↩
25. Dimaline, Marrow Thieves, Minerva’s song, as quoted and discussed in Gaertner, “Reconciliation.” ↩
26. Audra Simpson, “On Ethnographic Refusal: Indigeneity, ‘Voice’ and Colonial Citizenship,” Junctures, no. 9 (2007): 67–80, article. ↩
27. Donavyn Coffey, “Māori Are Trying to Save Their Language from Big Tech,” WIRED, April 28, 2021, report. ↩
28. Te Hiku Media, “Kaitiakitanga License,” preamble and terms 1–4, GitHub, accessed October 11, 2026, licence. ↩
29. Dimaline, Marrow Thieves, 230–31, quoted in Gaertner, “Reconciliation.” ↩
