From data sovereignty to AI capability

There is a tendency to discuss Māori technology futures at very large scale: sovereign clouds, Māori foundation models, national infrastructure and major programmes. Those ambitions may have value, but meaningful capability can also begin with much smaller and more practical steps.

The hardware and software threshold for local AI has fallen rapidly. Organisations can now experiment with open-weight models, local document retrieval and offline inference without building a research laboratory. The important opportunity is not to chase every new model. It is to deliberately build enough technical capability that Māori organisations can make their own choices about architecture, data and use.

Start with capability, not a grand platform

The first objective should be people who understand the technology. That means more than people who know how to write prompts. It means staff who understand what model inference is, how documents are retrieved, what embeddings do, how context reaches the model, what local processing changes, where logs are stored and how to test whether a system is behaving as expected.

This capability does not require sensitive Māori information. Public reports, open government data, published iwi management plans, synthetic datasets and GIS metadata are enough to learn almost all of the underlying techniques. That allows experimentation to begin without making the hardest governance decisions first.

Year one: learn and test

The first year should be primarily about learning. Install several local models. Compare their behaviour with commercial cloud models. Test local document retrieval. Try scanned reports, long PDFs, spreadsheets and GIS metadata. Test Māori names and macrons. Test questions where the answer is absent and see whether the model admits it does not know.

Record failures as carefully as successes. Does the model invent page references? Does it confuse similarly named places? Does OCR damage te reo Māori? Does a smaller model retrieve the right source but write a less polished answer? Does a larger model produce better prose while making the same factual mistake?

These experiments create a local evidence base. They also help organisations decide which capabilities genuinely matter. For a document-search system, accurate retrieval and source citation may matter far more than elegant writing.

Year two: build useful internal systems

The second year can move toward carefully bounded production uses. An iwi might create a local research assistant over approved environmental and planning material. A Māori organisation might build an internal search system for technical documentation. A GIS team could connect document retrieval with its metadata catalogue so staff can discover both spatial datasets and the reports that explain them.

This is also the point where access controls need to become more sophisticated. Not every collection should necessarily be available to every user, even within the same organisation. The system should be designed around the authority attached to the information rather than assuming that all internal information belongs in one giant index.

Evaluation should also become systematic. Build authorised test questions and expected source documents. Measure whether the system retrieves the right material, preserves names, cites the right pages and distinguishes evidence from inference. When a model is upgraded, run the same tests again. This turns AI from an impressive demonstration into a managed information system.

Year three: shared infrastructure where it makes sense

By the third year, some Māori organisations may be ready to operate or share more substantial AI infrastructure. Shared GPU servers could host several approved models while individual organisations retain control over their own information stores. Central technical expertise could maintain models and security without requiring every iwi or Māori organisation to build an identical engineering team.

This possibility becomes particularly interesting alongside Māori-controlled storage. Te Pā Tūwatawata demonstrates that Māori-owned distributed digital infrastructure is already practical. Compute could follow storage, allowing Māori-controlled data to be processed on Māori-controlled infrastructure while access remains determined by the organisations responsible for the information.

Shared infrastructure would need careful governance, security, funding and support. It should not mean pooling all Māori data into one central repository. Shared capability and shared compute can exist while data authority remains distributed.

Do not wait for a perfect Māori model

There is a risk of believing useful Māori AI must wait until a perfect Māori-specific foundation model exists. That could unnecessarily delay capability development. International open-weight models can be useful while their limits and provenance are treated seriously.

Their performance on te reo Māori and Māori contexts should be tested rather than assumed. Their training histories may be incomplete or contested. Their outputs should not be granted cultural authority. But none of this prevents them from performing bounded technical tasks such as finding every occurrence of a term across an authorised collection, extracting candidate metadata or helping explain a QGIS workflow.

Purpose matters. A model does not need authority over mātauranga to help a researcher locate the five pages in which a particular place name appears. It does need to return those pages accurately and make clear that the human researcher remains responsible for interpretation.

Build evaluation capability

One of the most important capabilities will be evaluation. Organisations should not have to rely entirely on global benchmark scores that say little about Aotearoa-specific work. A model that performs well on international reasoning tests may still mishandle Māori names, confuse local government terminology or struggle with scanned New Zealand historical documents.

Small local evaluation sets can be extremely useful. They can include questions with known answers, known source documents, names with macrons, ambiguous place names and examples where the right answer is that the information is not present. The goal is not to create a universal Māori AI benchmark overnight. It is to know whether a particular system is fit for the task it has been given.

Where appropriate authority exists, Māori organisations could also collaborate on non-sensitive evaluation material. This would allow different models and architectures to be compared without requiring organisations to share restricted information.

Develop local AI engineering skills

The technical skills worth building are increasingly clear. Organisations need people who understand local model installation, model selection, quantisation, memory and GPU requirements, RAG, embeddings, vector databases, document extraction, access control, model evaluation, logging, security and software supply chains.

They also need people who can integrate AI with the systems already used for real work. In Māori GIS that may mean QGIS, ArcGIS, spatial databases, field systems, document repositories and metadata catalogues. In other organisations it may mean records systems, environmental monitoring platforms, policy repositories or research archives.

The useful AI is often not another chatbot. It is a capability embedded into an existing workflow where it removes repetitive effort while preserving the source and the authority of the people doing the work.

Keep humans where authority sits

As AI becomes more capable, the boundary between assistance and authority should become clearer rather than blurrier. Automate effort where appropriate. Do not automate authority simply because it is technically possible.

Use AI to search, organise, compare, extract, draft and expose forgotten information. Let it help people work across collections too large to read manually. But where decisions depend on tikanga, whakapapa, local authority, contested history or significant consequences for people, the system should support those responsible for the decision rather than quietly replacing them.

This is not technological conservatism. It is good system design. The strongest AI systems are often those that are explicit about what the model is responsible for and what remains a human responsibility.

Avoid two bad futures

There are two unhelpful futures. In the first, enthusiasm wins and Māori information is poured into every new AI service because the tools are useful and convenient. Dependency grows, context is lost and decisions about infrastructure are made largely by overseas vendors.

In the second, fear wins and AI becomes so tightly associated with risk that practical experimentation slows. Technical expertise develops elsewhere, Māori organisations remain consumers of commercial platforms, and the ability to build alternatives never matures.

A stronger path sits between these. Use governance to make experimentation safe enough to learn. Use experimentation to make governance technically informed. Build local capability so that organisations have more options rather than fewer.

A positive direction

Māori data sovereignty should not leave Māori organisations standing outside artificial intelligence asking whether they are permitted to enter. Nor should enthusiasm for AI sweep aside hard-won principles about control, whakapapa, collective benefit and kaitiakitanga.

The more interesting path is to use those principles to shape the technology. Choose where information lives. Choose which models process it. Choose what leaves the environment. Choose who has access. Choose which outputs require human verification. Choose what information should never enter an AI system. Build people who understand enough technology to know whether those choices are actually being implemented.

Where existing commercial systems provide sufficient control, use them. Where they do not, build the capability to operate alternatives. Where AI adds little value, do not use it merely because it is fashionable.

The future of Māori AI does not need to be framed primarily as resistance to technology. It can be framed as increasing agency over technology. The important question is no longer only how Māori data can be protected from AI. It is increasingly how Māori can have the knowledge, infrastructure and authority to decide what AI will do with Māori data.

That shift from protection alone toward protection plus capability is the central recommendation of this series.

Return to the beginning of the series with Where does your data actually go? Six ways to use AI.