As governments and companies race to fold artificial intelligence into everything from health records to education, a quieter argument is gathering force in Australian research circles: the systems being built at speed are drawing on Indigenous Knowledges without the consent, context or control of the communities those knowledges belong to. Researchers at the University of Melbourne have put that concern front and centre, arguing that AI must be built with Indigenous Knowledges rather than against them.
The framing matters because it inverts how most of the current AI conversation is conducted. The dominant story is about scale, compute and the race to deploy, with cultural and ethical questions treated as compliance problems to be tidied up later. The Melbourne position, set out in the University of Melbourne’s analysis, treats those questions as foundational. If Indigenous data, languages, stories and ways of knowing are going to shape how machines reason about the world, the argument runs, then First Nations people need to be designers and decision-makers in that process, not subjects of it.
The context: extraction by default
Large AI models learn by ingesting vast quantities of text, images and audio scraped from the open web and from digitised archives. That method is indiscriminate by design, and it captures Indigenous cultural material along with everything else. Language recordings, artwork, oral histories and ceremonial knowledge that were shared in specific contexts, sometimes with strict cultural protocols about who may see or hear them, can end up flattened into training data and reproduced by a chatbot with no understanding of their significance.
This is where the concept of Indigenous data sovereignty enters the picture. The idea, developed over the past decade through networks such as the Maiam nayri Wingara collective in Australia, holds that Indigenous peoples have the right to govern the collection, ownership and use of data about their communities, lands and cultures. Applied to AI, it raises hard questions that the technology industry has largely avoided. Who consented to this material being used? Who benefits when a model trained on it is sold? And what happens when a system confidently generates false or culturally inappropriate content while claiming to speak for a community?
The news: a call to build differently
The Melbourne argument is less a single announcement than a statement of principle aimed at researchers, developers and policymakers. Its core claim is that Indigenous Knowledges are not simply another dataset to be mined, but living systems of understanding that carry their own governance, relationships and responsibilities. Building AI “with” those knowledges means genuine partnership: co-design, community control over what is included and excluded, and a share in whatever value the resulting tools create.
Crucially, the argument is not anti-technology. It points toward a more constructive possibility, in which AI could support the revival of endangered languages, help manage Country using both traditional ecological knowledge and modern sensing, or preserve cultural material on terms set by the communities themselves. The difference is who holds the steering wheel. Tools built on extraction tend to reproduce the power imbalances that produced them. Tools built on partnership can, at least in principle, do the opposite.
Two ways of seeing the problem
Not everyone frames the challenge the same way. For many technologists and firms, the pressing task is practical mitigation: better dataset documentation, opt-out mechanisms, filters that catch sensitive material, and clearer licensing. On this view, the harms are real but manageable within the existing development pipeline, and slowing down to renegotiate the foundations would simply hand the advantage to overseas models that observe no such scruples.
The Indigenous data sovereignty position pushes back on that logic. It argues that bolting protections onto a system designed around extraction will always leave communities in a defensive crouch, reacting to decisions made without them. The more durable answer, on this reading, is structural: embed First Nations governance at the point where data is gathered and models are designed, so that consent and benefit are built in rather than patched on. Both camps agree the status quo is inadequate. They disagree about whether the fix is a feature or a foundation.
The Australian stakes
For Australia, this is not an abstract ethics seminar. The country is home to the oldest continuous cultures on Earth and to hundreds of Indigenous languages, many of them endangered, all of them increasingly digitised. As Australian universities, agencies such as CSIRO and a growing crop of local firms build sovereign models and data infrastructure, the choices they make about training material will determine whether Indigenous Knowledges are protected or quietly absorbed. The federal government’s evolving AI framework, its push for guardrails around high-risk uses, and its interest in a distinctly Australian AI capability all intersect with this question.
There is also a competitive dimension that sits oddly well with the ethical one. Australia has argued that its edge in AI will come not from matching American compute budgets but from doing specialised, trustworthy, locally grounded work. Indigenous data sovereignty is a natural test of that ambition. Getting it right would demonstrate that respectful, consent-based AI is achievable at scale, and it would give Australian institutions genuine expertise to export. Getting it wrong would repeat a long history of cultural material being taken without permission, this time at machine speed and global reach.
What’s next
The practical path forward is already visible in outline. It runs through community-controlled data protocols, formal agreements between researchers and Traditional Owners, and funding models that direct value back to the communities whose knowledge makes the tools possible. It also depends on Indigenous people being trained and employed as AI practitioners, so that partnership is real rather than symbolic. None of this is simple, and it will move more slowly than the technology it is trying to shape.
The harder task is cultural. It asks an industry built on speed and scale to accept limits it did not set, and to treat consent as a design constraint rather than an obstacle. Whether Australia’s AI sector proves willing to build with Indigenous Knowledges, rather than around them, will say a great deal about the kind of AI nation the country intends to become.
Sources: The University of Melbourne.
