When we say that Artificial Intelligence can assist with legal research, legal reasoning, or the production of legal knowledge, a prior question deserves far more attention than it receives: whose law do these systems know?

Large language models acquire their legal competence from text: judgments, statutes, treaties, journal articles, and regulatory instruments that have been digitised, indexed, and made computationally accessible at scale. The composition of these corpora is neither neutral nor complete. As Bender et al. observed in their critique,¹⁾ training data encode the hegemonies of the archives from which they are drawn; they overrepresent dominant languages, institutions, and viewpoints, and they exclude communities whose knowledge circulates outside mainstream publishing infrastructures.

The legal traditions of Indigenous Peoples and of much of the non-Western world are characteristically oral, relational, etched in carvings, seasonal, performative, and place-based. Their normative content lives in narrative, genealogy, ceremony, and the deliberations of communities; their authority is constituted through relationships between people, land, ancestors, and the non-human world. An AI system built to learn from text does not merely fail to locate such law. It is architecturally incapable of recognising legal knowledge in the forms in which that law exists.

I write from New Zealand, where this problem is neither abstract nor new. It is the digital continuation of a constitutional argument that Māori (the Indigenous Peoples of New Zealand) have been making for nearly two centuries: that the question of which knowledge counts as law is, at bottom, a question of who holds authority, and that colonial institutions have always answered it in their own favour. What has changed is the infrastructure. The determination of what counts as legal knowledge is now being embedded, silently and at planetary scale, into systems that most lawyers will use without ever asking what those systems were built to see.

A Living Jurisprudence

Tikanga (Māori Legal Knowledge), involving the system of values, principles, obligations, and dispute-resolution processes developed by Māori over centuries, offers an instructive case for readers accustomed to thinking about legal pluralism in constitutional terms. Tikanga is not folklore adjacent to “real” law. New Zealand’s Supreme Court accepted in Ellis v R (2022) that tikanga was the first law of Aotearoa and confirmed it remains a source of New Zealand law, capable of application beyond disputes involving only Māori parties.²⁾

The Law Commission’s landmark study paper He Poutama (2023) has since mapped the deepening interweaving of tikanga and state law in what Justice Joe Williams has memorably called an emerging lex Aotearoa.³⁾ For comparative constitutional lawyers, this is recognisable terrain: a settler state grappling, belatedly and imperfectly, with the co-existence of two legal orders within one constitutional space.

But recognition has proceeded through codification, the translation of a living jurisprudence into textual forms legible to courts, tribunals, and legislatures. And codification is where the epistemic risk begins. Tikanga is not a fixed body of doctrine. It is a deliberative practice, exercised through discussion and collective deliberation, weighted by genealogy and relational standing, and applied contextually by those holding recognised authority within the tribal and sub-tribal communities in whom that authority vests. What enters the written archive (digitised legislation, tribunal reports, judgments, statutory definitions, academic commentary) is a filtered representation, shaped by the evidentiary demands of the forums for which it was produced.

AI systems almost exclusively ingest only this filtered layer and then filter it again through the statistical logic of the model. The result is a second-order artefact: a reconstruction of state and academic interpretations of tikanga, detached from the communities whose living practice gives tikanga its authority. A researcher anywhere in the world who asks a legal AI about Māori law will receive fluent, confident text. What they will not receive, and will not know they are not receiving, is any trace of the deliberative processes, internal authorities, and relational obligations through which Māori communities themselves determine what tikanga requires. Fluency without standing is not legal knowledge. It is presented with the authority of apparent comprehensiveness, while functioning as a vector for misinformation.

The Structural Problem

The structural problem generalises well beyond New Zealand, and three features of Indigenous and non-Western legal epistemologies show why.

Orality

Many Indigenous legal orders, in New Zealand, Australia, across the Pacific, and throughout the Americas, transmit normative content through oral tradition, where accuracy and authority are secured by relational and ceremonial protocols rather than by fixation in writing. John Borrows has shown, for the Canadian context, that these are sophisticated constitutional orders in their own right. Oral legal knowledge that has never been transcribed simply does not exist for a text-trained model; oral knowledge that has been transcribed exists only in the mediated form given to it by the transcriber, frequently a colonial ethnographer whose interpretive frame is embedded in the record itself.

Restriction

In many traditions, legal knowledge is deliberately confined to designated knowledge holders and disclosed only in appropriate contexts. This is not underdevelopment awaiting digitisation; it is a governance mechanism, in which the integrity of knowledge depends on proper custodianship. The corpus-building ethos of contemporary AI, in which comprehensiveness is treated as an unambiguous virtue, is in direct normative conflict with legal orders for which selective transmission is itself a legal principle.

Relational epistemology

In many Indigenous frameworks, legal knowledge is not a body of rules applied to facts but a property of relationships: between knower and known, between authority and community, between generations. Who may speak on a legal question is itself a legal question, answered by genealogy and community standing rather than institutional credential. Comparable structures run through other non-Western traditions: classical Islamic jurisprudence locates authority in chains of interpretive transmission and scholarly licence rather than in text alone; Ubuntu jurisprudence in Southern Africa grounds normativity in the relational constitution of personhood. Comparative law has long struggled with its own marginalisation of this plurality. AI now automates that marginalisation, accelerates it, and renders it invisible behind an interface.

From Invisibility to Extraction

It would be a mistake, however, to frame this only as a problem of absence. Where Indigenous knowledge is present in training corpora (digitised tribunal evidence, ethnographic records, documentation of Māori knowledge systems), a different harm operates – extraction. AI systems reproduce that knowledge, often without the consent of the communities from whom it derives, without any mechanism for those communities to govern its downstream use, and without recognition that, within the source legal order, such knowledge may be owned, restricted, or sacred.

The Indigenous Data Sovereignty movement has spent over a decade establishing that Indigenous Peoples hold inherent rights to govern data about their peoples, territories, and knowledge systems. In New Zealand, that claim carries constitutional force. In Te Tiriti o Waitangi, the 1840 treaty between the British Crown and Māori chiefs and a foundational instrument of New Zealand’s unwritten constitution, Article 2 of the Māori text guarantees tino rangatiratanga, commonly understood by the Waitangi Tribunal and many Māori scholars as affirming Māori authority over their lands, communities, and taonga (a term encompassing all treasured things, tangible and intangible). The Waitangi Tribunal,⁴⁾ the standing commission of inquiry into Crown breaches of the Treaty, held in its Ko Aotearoa Tēnei report (2011) that the Crown’s obligations extend to Māori knowledge.⁵⁾ In its WAI 2522 inquiry into the CPTPP trade agreement (2021),⁶⁾ in which this author gave expert evidence, the Tribunal recognised that data can record Māori knowledge and can itself be a taonga. It found the Crown had breached Treaty obligations in aspects of consultation and active protection relating to Māori interests in the digital environment.

The author has argued elsewhere that these guarantees generate specific obligations for AI: Māori data are taonga; their collection and use require Māori governance; and Te Tiriti operates as a constitutional threshold for the deployment of such technologies, not a consultative checkpoint to be satisfied after the design decisions have been made.⁷⁾

On this analysis, the ingestion of Māori legal and cultural materials into training corpora without the authority of Indigenous Peoples is not an intellectual property irregularity. On this account, it is a Treaty breach: the same extractive logic that alienated land and suppressed the Māori language and customs, as it did to other colonised Indigenous Peoples, now replicated in computational form. And a critical corollary follows for anyone tempted to treat the invisibility problem as a coverage problem. Digitisation programmes that pursue more inclusive corpora while leaving governance in the hands of technology firms and state archives do not remedy the harm; they compound it. Visibility on the wrong terms is but a new enclosure.

Four Responsibilities

What follows for legal scholars and institutions, I suggest, are four responsibilities, and none has yet been met.

The first is epistemic: to cultivate, and to teach, the capacity to perceive what AI cannot. Legal education must equip researchers to read a court decision as a mediated representation of Indigenous law rather than as the law itself, and to know where living legal authority actually resides; in which communities, held by which knowledge keepers, exercised through which deliberative processes. A generation trained to treat the model’s output as the horizon of legal knowledge will not know what it is missing and will not be trained to ask.

The second is structural: to recognise that decisions about what is digitised, indexed, and licensed for training are decisions about what AI will know. Law faculties, publishers, and archival institutions are not bystanders to corpus construction; they are its infrastructure. Where their choices systematically exclude Indigenous and non-Western materials, or include them without authority, they embed epistemic hierarchy into the foundations of future legal knowledge production.

The third is relational: institutions developing or deploying AI in legal contexts owe obligations to the Indigenous communities whose legal systems are implicated: obligations of governance, not merely consultation. In New Zealand this is a Treaty requirement: tribes and subtribes participate as rights holders exercising authority, not as stakeholders invited to comment on decisions taken elsewhere. Elsewhere, analogous duties flow from the UN Declaration on the Rights of Indigenous Peoples⁸⁾ and its standard of free, prior, and informed consent.

The fourth is critical: to resist the normalisation of AI’s limitations as though they marked the boundaries of law itself. The deepest risk here is not defective output but epistemic capture: that sustained reliance on AI-mediated research quietly trains the profession to treat the codified, the archived, and the institutionally recognised as the whole of law, and everything beyond that frame as pre-legal, customary in the pejorative sense, or merely cultural. Comparative lawyers will recognise this as an old violence in a new medium. Its novelty lies in its speed, its scale, and its invisibility.

Law as Relationship

Legal knowledge, in the traditions I have described, is not primarily a property of texts. It is a property of relationships: between communities and their histories, between knowledge holders and their obligations, between law and the land from which it emerges. AI systems as presently constituted cannot hold relationships. They hold textual representations of relationships, filtered through the archival practices of the institutions that have historically dominated legal publishing.

That constraint will not be dissolved by scale, by multilingual corpora, or by better retrieval. It demands a reckoning with the epistemological foundations of these systems: what they are designed to know, and whose knowledge they are structured to recognise. That reckoning is, at its core, a constitutional question, and like all constitutional questions in a Treaty-based order, it cannot be answered without first asking who holds the authority to speak. For legal scholarship, engaging that question is not optional. To proceed as though AI’s blind spots were merely technical is itself a jurisprudential position, and it is one that the communities whose rights depend on legal recognition are entitled to hold us to account for.