Indigenous data sovereignty and the authority to determine what knowledge becomes in AI systems
In June 2024, UBC Arts published “Indigenous data stewardship stands against extractivist AI,” an interview written and framed by Ritwik Bhattacharjee. In it, I argued that AI’s appetite for knowledge had to be understood through Indigenous data sovereignty and colonial extraction, and I tried to draw attention away from the apparent magic of the output and toward the relationships its production disregards. Returning to it, I want to follow that argument further. Collection begins a relationship whose consequences extend through what a system retrieves, how it summarizes, what it infers, and where its answers circulate. An acknowledgment of the source can survive all these operations while the source community’s authority is excluded from decisions about its use.
The interview also carried a tension. One question asked what Indigenous adaptation to past digital technologies could teach us about AI. I contested the adaptation story, but the framing still risked making Indigenous stewardship a resource for developers: useful principles travel into the industry while the authority to determine their application stays behind. Indigenous authority over knowledge cannot depend on whether respecting it makes our technologies better. It may require developers, researchers, and universities to relinquish access, change a project’s purposes, or accept that a proposed use should not proceed.
A story’s availability tells us that someone can encounter it. It does not tell us who decided it should be available, or what the encounter entitles anyone to do. Once computational systems begin working on knowledge, who determines what it becomes? By authority I mean jurisdiction over organized uses, not control over every interpretation. A community does not need to fix what a story means in order to govern the services, infrastructures, and purposes through which it is made to circulate.
Access and authority
Local Contexts’ Traditional Knowledge Non-Commercial Label communicates a condition on the use of cultural material: access does not authorize commercial use. Communities customize the label to express their own requirements. It addresses someone who can encounter the material and nevertheless remains responsible for how they use it. [1]
My previous essay, “Manufactured Indigenous Belonging: What an AI Label Can’t Authorize,” looked at a label that discloses how We-News’s AI reporters are made while leaving their claims of Indigenous belonging standing. The problem here runs the other way. This label carries a community’s own authority. The question is whether that authority governs what a system does with the material.
Consider a hypothetical. A developer builds a commercial research assistant that retrieves passages from an online collection and supplies them to a language model, which composes an answer from them. This is retrieval, distinct from training a model on the collection and from saving conversational memories or profiling users, operations at the centre of “Mining Interiority.” The collection includes material carrying the TK Non-Commercial Label. The assistant produces a fluent answer, perhaps even a citation. Has the condition on use survived the passage from collection to answer?
The example requires little technical drama. The source need not disappear, and the answer need not be inaccurate. A system could preserve attribution while the authority behind the label has no influence over the service built on it: the label stays perfectly legible and governs nothing that happens next. Greater accuracy could make the unauthorized use more effective.
Openness describes an arrangement while obscuring who arranged it. An archive makes a document readable; a researcher reproduces it; a developer retrieves it for a service. Each action has its own purpose and incurs its own responsibilities. Permission for one does not settle authorization for the next. Publication also leaves open who was entitled to authorize circulation in the first place. An institution may hold a record without holding authority over it, and making that record easier to find extends the institution’s power if no one asks how it acquired the power to decide.
Copyright addresses some disputes over reproduction and use. Its categories do not exhaust the collective responsibilities, relationships to territory, and community protocols that govern knowledge. A use can be lawful under copyright and still disregard those obligations.
A system’s capacity to interpret knowledge must also be distinguished from its authority to use it. Interpretation, however faithful, does not settle jurisdiction. Permission asks whether an activity has been authorized. Jurisdiction asks who establishes the conditions under which authorization can occur, who interprets those conditions, and whose decisions have force when a dispute arises. If a company retains the final power to determine which community requirements are practical enough to honour, consultation has left decisive control with the company.
That is the operation I mean by extraction. Material enters a system whose operators acquire control over its subsequent uses, while the people responsible for it are denied effective control over those uses. The colonial relationship lies in that displacement of existing authority and in the concentration of power over what follows.
Technological histories with obligations
The assumption that Indigenous peoples merely adapt to technologies invented elsewhere already assigns the power to define technological futures. It makes participation look like catching up. In 2024, I contested that assumption with an argument from precedence, describing wampum belts as practices of data stewardship and hypertextuality that came long before the digital age gave those concepts their names. Precedence matters, but it is not enough.
Angela M. Haas makes the historical case in “Wampum as Hypertext.” Her account of wampum, including its diplomatic use among Haudenosaunee nations, connects material design with memory, performance, and relations among parties. Wampum records require practices of revisiting and reading; accepting wampum can entail reciprocal responsibilities. [2]
The historical argument contests the story that technological invention belongs to the West. Stopping there, however, leaves that story’s standards of achievement intact: Indigenous technologies receive recognition for anticipating categories through which Western institutions already recognize themselves. My earlier framing left the implications of those reciprocal responsibilities largely undeveloped.
Haas’s account supports a further question about the relation between a record and its use. If receiving a communication entails responsibility, accessibility cannot be evidence that the communication has shed its obligations. A record can travel without its recipient becoming entitled to establish its purpose.
I draw this argument from specific, living intellectual and diplomatic practices. Wampum traditions carry authorities and responsibilities that cannot be detached from them and repackaged as general principles for AI development. They challenge the assumption that technical sophistication requires knowledge to circulate independently of those authorized to interpret and care for it.
That challenge also changes what stewardship can mean. In the interview, stewardship sometimes appeared as a way to improve the treatment of material inside technological systems, as though its eventual incorporation into a dataset, model, or product were already settled. Sovereignty puts that incorporation itself in question.
Deloria and the institutions of knowing
Vine Deloria Jr.’s The Right to Know, prepared for a 1978 White House pre-conference on Indian library and information services, brings the institutional conditions of knowledge into view. In David E. Wilkins’s account, Deloria called for direct federal funding of tribal library and archival services and the transfer of records to facilities on reservations. Wilkins also emphasizes the trained staff and equipment these institutions require. Knowing depends on resources and institutions through which Native nations can sustain their intellectual and political lives. [3]
That specificity matters. An institution can celebrate Indigenous knowledge while retaining the money, infrastructure, and decision-making power that determine its use. Glen Coulthard’s critique of the colonial politics of recognition explains why the arrangement is so durable: recognition can be extended while the relations of dispossession it claims to address stay in place. [4] Recognition makes few demands on an institution willing to acknowledge authority but unwilling to relinquish control.
I read Deloria’s intervention as a challenge to any account of access that leaves those arrangements untouched. His argument concerned archives, education, and the responsibilities of governments. Extending it to AI means attending to the institutions that now organize computational access: who operates the service, who funds its maintenance, and who can alter its purposes.
Indigenous data-governance frameworks already insist on this. The First Nations principles of OCAP®, established in 1998 and stewarded by the First Nations Information Governance Centre, set out ownership, control, access, and possession as distinct dimensions of First Nations authority over their information. Access includes the right of First Nations to make decisions about others’ access to their collective information; it is one dimension of authority among four, not a stand-in for the whole. FNIGC is explicit that these are First Nations principles rather than pan-Indigenous ones. [5] The Global Indigenous Data Alliance’s CARE Principles for Indigenous Data Governance place collective benefit alongside authority to control, responsibility, and ethics. They bring people and purposes into data governance alongside the FAIR goals of making data findable, accessible, interoperable, and reusable. [6]
Sovereignty also includes the authority to build, commission, and sustain technological projects. A community might seek wider circulation, commercial collaboration, restricted access, or a project that outside developers would never have proposed. Respecting sovereignty means that these purposes cannot be reduced to whichever option best fits an existing product.
SIKU offers a concrete example. Built and operated by the Arctic Eider Society, an Indigenous-driven charity based in Sanikiluaq, Nunavut, the platform supports Indigenous-led research, monitoring, and knowledge transfer. Its published governance distinguishes access for particular purposes from permission for further uses. Public posts can be read for safety and learning, but visibility does not authorize other users to export or manipulate them. Research use requires consent through both a written project licence and the contributor’s choice of which posts to share. Projects must also meet applicable community and regional requirements for licensing, ethics review, and approval. The Arctic Eider Society must seek permission to reuse posts in the same way as other users. [7]
SIKU’s documented arrangements show how a platform can distinguish purposes, recipients, and permissions within digital circulation. The principle carries into AI retrieval: access need not flatten those distinctions, and operating the infrastructure need not confer a general entitlement to reuse what it holds.
Refusal belongs to this authority to determine what happens. In Mohawk Interruptus, Audra Simpson develops a politics of refusal through the sovereignty of Kahnawà:ke Mohawks, challenging the demand that Indigenous political life seek its validation through settler recognition. Eve Tuck and K. Wayne Yang describe refusal in research as generative: it marks what the academy is not entitled to know and turns inquiry toward the power relations that make knowledge available. [8] This capacity becomes consequential precisely when a community’s decision conflicts with a developer’s intentions.
What survives transformation
Return to the hypothetical research assistant. Its operators can restrict which collections enter retrieval and establish procedures through which cultural authorities determine permissible uses. A citation demonstrates neither that those arrangements exist nor that they govern the service. The relevant question is what the operator is required to do when an intended use exceeds its authorization.
Local Contexts distinguishes among labels that communicate provenance, labels that articulate protocols, and labels that indicate permissions. Uses beyond those permissions require engagement with the relevant cultural authorities. [9] The distinction matters for AI because a source reference can acknowledge where material came from while doing little to constrain where it goes.
What I have begun to call a computational afterlife concerns these subsequent transformations: what systems make from material, and what uses their outputs enable. Walter Benjamin described a work’s afterlife in translation as the transformation and renewal of something living. [10] I borrow the term to hold open a tension. Computational transformation sustains circulation, but circulation can continue while the relationships that make knowledge responsible are left behind. An output’s continuing usefulness is no assurance that those relationships survive. Its afterlife remains a question of jurisdiction.
Follow the hypothetical one step further. The assistant’s answer, with its citation, is pasted into a consultancy’s report on a proposed development, and a client acts on the report. The community whose knowledge the answer drew on learns of the use and contests it. Suppose the operator accepts that the use was unauthorized. What does a remedy reach? Removing the material from retrieval changes future answers. It does not recall the report or reverse the client’s decision. Responsibilities now extend across institutions, and responding may require notifying recipients, withdrawing the report, and revisiting decisions made through it. Those actions cannot undo every consequence, but an obligation does not expire because the material has moved beyond the service that first processed it.
If transformation released outputs from the obligations attached to their sources, every new output would become an occasion to declare those obligations void. The ability to transform knowledge would become the power to decide when responsibility ends.
The difficulty is therefore larger than whether an answer retains enough context. Context can itself become a product feature: richer provenance, fuller description, more culturally attentive phrasing. A system can describe the obligations surrounding knowledge while its operator remains unanswerable to the people responsible for it. The description may be accurate even as the service disregards everything it says the relationship requires. Provenance alone leaves open whether those obligations change what the operator must do.
There is a parallel problem with individual users: a person may agree to one interaction without knowing which later inferences or uses will follow. Indigenous collective sovereignty raises a distinct question that individual consent cannot settle. SIKU foregrounds contributors’ ownership of their posts while also requiring respect for community governance. Both matter. Permission from one contributor cannot substitute for the collective authorization a proposed use requires. A community member’s upload, however well intentioned, is not by itself a community’s decision.
Making jurisdiction consequential
Indigenous jurisdiction does not originate with the operator. What a governance arrangement can change is the distribution of effective control, resources, and enforceable obligations, so that jurisdiction has consequences inside the service rather than only in the service’s account of its sources. This is a material arrangement as much as a procedural one. A commercial service can earn revenue from an answer while a community bears the expense of discovering how its knowledge was used and seeking redress.
A governance agreement worth the name would give designated community authorities the means to investigate contested uses, including access to relevant records of what the system retrieved and produced. It would establish their power to require changes or suspend an authorized use without having to secure the operator’s agreement anew. It would also assign responsibility for responding when an output has already travelled into another institution’s work.
These powers require resources. An agreement must establish who pays for monitoring and disputes, who interprets contested terms, and where a community can require a response when the operator refuses one. Otherwise, the institution benefiting from the use can also determine whether a challenge to it will have consequences. Consultation that leaves that arrangement intact has changed the conversation while leaving effective control where it was.
Competence and relinquishment
In the original interview, I described teaching about power and technology as a matter of professional competence. I would now make the obligation more concrete. Before asking students whether an AI system gives a useful answer, I want them to ask what they are authorized to supply to it and what the system is authorized to do with that material.
A student may possess a document without having authority to upload it to an external service. A university may approve a platform without settling the permissions governing every collection used within it. Technical availability relieves neither of the responsibility to investigate.
As a settler scholar working in Indigenous literary studies, I am bound by the same demands, as is the university that employs me. A persuasive critique creates no exemption from the arrangements it describes. The obligation is to make those arrangements answerable to the people whose knowledge enables them, including when their decisions require abandoning an intended use.
The question I carry forward from 2024 is what AI development would have to relinquish to respect Indigenous jurisdiction. Sometimes it would be access to particular material. Sometimes it would be unilateral control over a project’s purposes. It might require funding infrastructure whose value cannot be measured by its contribution to the developer’s service.
Knowledge held in stories, records, and wampum lives through particular places and relationships. Making it computationally useful does not end those relationships. The label remains an address to someone capable of responding. A model can generate an acknowledgment; the response the label asks for belongs to the operator, who can alter the service, honour the restriction, hand over records, or stop a use. What would require that institution, and the universities adopting its products, to honour the obligations it can already describe?
CONTINUE ──
A Dream Is Not a Resource approaches the authority to refuse through Cherie Dimaline’s novel and the relationships personalized AI turns into resources.
Notes
1. Local Contexts, “TK Non-Commercial (TK NC).” The research assistant is a hypothetical application of the documented label, rather than an allegation about a particular service. ↩
2. Angela M. Haas, “Wampum as Hypertext: An American Indian Intellectual Tradition of Multimedia Theory and Practice,” Studies in American Indian Literatures 19, no. 4 (Winter 2007): 77–100. The article’s opening discussion connects the reading and maintenance of wampum with reciprocity and responsibility. ↩
3. Vine Deloria Jr., The Right to Know, prepared for the White House Pre-Conference on Indian Library and Information Services On or Near Reservations, Denver, October 1978. The account here follows David E. Wilkins, “The Right to Know: Another Deloria Suggestion,” Indian Country Today, July 26, 2009. Wilkins quotes Deloria’s proposals for federal funding and records transfer and discusses the staffing and equipment such institutions require. ↩
4. Glen Sean Coulthard, Red Skin, White Masks: Rejecting the Colonial Politics of Recognition (Minneapolis: University of Minnesota Press, 2014). ↩
5. First Nations Information Governance Centre, “The First Nations Principles of OCAP®.” OCAP® is a registered trademark of FNIGC. ↩
6. Stephanie Russo Carroll et al., “The CARE Principles for Indigenous Data Governance,” Data Science Journal 19 (2020): article 43. See also the Global Indigenous Data Alliance’s CARE Principles overview. ↩
7. SIKU, “Indigenous Data Sovereignty” and “Guiding Principles” and “Terms of Use” (terms updated April 13, 2026); SIKU Support, “Understanding Project Stewardship” and “Post Stewardship Settings.” These sources document the platform’s governance commitments. The essay draws an argument from those arrangements; it does not independently assess their implementation or report on an AI retrieval system. ↩
8. Audra Simpson, Mohawk Interruptus: Political Life Across the Borders of Settler States (Durham, NC: Duke University Press, 2014); Eve Tuck and K. Wayne Yang, “R-Words: Refusing Research,” in Humanizing Research: Decolonizing Qualitative Inquiry with Youth and Communities, ed. Django Paris and Maisha T. Winn (Thousand Oaks, CA: SAGE, 2014), 223–47, especially 241–42. ↩
9. Local Contexts, “TK Labels,” particularly its distinction among provenance, protocol, and permission labels. ↩
10. Walter Benjamin, “The Task of the Translator,” in Illuminations, ed. Hannah Arendt, trans. Harry Zohn (New York: Schocken, 1968), 69–82. ↩
