A better way to assess Māori AI risk

AI governance often ends with a rating such as low, medium or high risk. Those labels can be useful, but only after somebody understands what the system actually does. Rating “AI” in the abstract is much less useful than examining the information, the architecture, the authority and the consequence of a particular use.

A more practical Māori AI assessment can begin with four dimensions. They do not replace tikanga, organisational policy or legal obligations. They help make the technical system visible enough for those things to be applied proportionately.

What information is involved?

The first dimension is the information itself. Public environmental data creates a different issue from whakapapa. A published Waitangi Tribunal report creates a different issue from confidential negotiation material. A general GIS metadata catalogue creates a different issue from precise coordinates of sensitive sites.

There is no meaningful AI risk assessment without understanding the material. The fact that a document is digital, that somebody can technically open it, or that it has already been published somewhere does not automatically settle whether combining it with other information or making it instantly searchable is appropriate.

This is where Māori data-sovereignty principles remain essential. Te Mana Raraunga treats Māori data as connected to people and collective interests rather than merely as records detached from their origins. The question is therefore not only whether an organisation legally possesses information, but what authority and responsibility travel with it.

What architecture is being used?

The second dimension is where processing occurs. Does the information enter a public chatbot, an enterprise cloud service or a commercial API? Are the documents kept locally while retrieved passages are sent to a cloud model? Do the embeddings remain local? Does the model itself run locally? Can the system operate with its internet connection disabled?

Architecture can remove entire categories of risk. If a sensitive document never leaves an approved local environment, the risk of an overseas AI provider retaining that document may disappear or materially reduce. Other risks remain, such as internal access, inappropriate interpretation and model quality.

Good governance should recognise both sides of that equation. A control that makes sense for a public chatbot may be unnecessary for a genuinely offline local model, while an entirely local system may still deserve very strong controls if it contains highly sensitive information.

Who has authority?

The third dimension is authority. This is where conventional privacy or cyber-security assessment can be too narrow. A person may have technical access to a document while collective interests, whakapapa relationships or the purpose for which information was originally gathered still matter.

AI can also change the practical meaning of access. Giving somebody read access to 20,000 historical documents is not quite the same capability as giving that person an AI system that can immediately extract every reference to a named whānau, place or event. The underlying files may be unchanged, but the effective power to discover and connect information has increased.

Authority should therefore be considered at the level of collections and uses, not only individual files. Some material may be suitable for ordinary staff search, some for a specific project group, some for people with particular authority, and some should not be indexed at all.

What happens because of the output?

The fourth dimension is consequence. An AI system helping a researcher find relevant documents is one thing. An AI system deciding which families warrant investigation, influencing eligibility for services or determining health priorities is another.

A summarisation error may waste time. An incorrect decision in policing, welfare, health or education can materially affect a person's life. Consequential systems should therefore receive strong scrutiny even if every byte remains inside Aotearoa and the infrastructure is Māori owned.

Local hosting cannot compensate for a bad decision model. Likewise, a commercially hosted tool may be entirely reasonable for a low-consequence public-information task. Consequence needs to be assessed independently from data location.

Put the four together

These four dimensions produce a more proportionate view. Public information, authorised users, local processing and low-consequence research may require relatively light controls. Highly sensitive information, unclear authority, external processing and decisions affecting people demand much stronger governance.

Between those extremes are many legitimate uses. The purpose of governance should be helping people navigate that middle ground rather than forcing every AI use into the same category.

A local GIS metadata assistant, for example, might contain only approved technical information and help staff locate datasets. Its risk profile could be modest. A local model searching sensitive whakapapa might have excellent infrastructure controls but still require very strong authority and access arrangements. A government cloud system predicting risk about Māori communities may need intensive scrutiny even if the provider offers excellent enterprise privacy protections.

Replace slogans with system statements

Every proposed Māori AI use should have a short architecture-and-consequence statement. It might say: “This tool searches publicly available environmental reports. All documents, embeddings and model inference remain on an iwi-owned workstation. The output is used only to help researchers locate source documents and is not treated as authoritative.”

Another might say: “This system sends selected internal documents to a commercial model under an enterprise agreement. The outputs are used to prepare draft policy analysis. Restricted information categories are excluded, prompts are logged under the enterprise service, and all substantive conclusions require human verification.”

Those paragraphs tell decision-makers considerably more than a generic declaration that an organisation uses responsible AI. They also make it possible for technical staff to prove whether the system actually operates as described.

Separate model risk from data-flow risk

A useful assessment should also distinguish questions about a model's original training from questions about current organisational data. An open-weight model running locally may have unclear training provenance or poor representation of Māori contexts. Those are real concerns. They are different from the question of whether a current iwi document is transmitted to an overseas service.

Separating the two makes governance clearer. The organisation can evaluate model provenance, licensing, bias and performance while separately verifying the data path, logging, storage and retention of its own material. One problem does not need to be exaggerated to make the other important.

Test rather than assume

Many AI risks can be tested. Disconnect the network and see whether the local application still functions. Inspect logs. Ask whether embeddings are stored on disk. Test permissions with accounts that should not see a collection. Try to make the model reveal a source it should not have access to. Check whether citations lead to the correct page. Compare provider documentation with observed behaviour.

This practical testing is important because AI governance should not depend entirely on marketing claims or generic policy language. Technical assurance is part of responsible governance.

The same applies to performance. Build small evaluation sets using authorised material. Measure whether the model finds the right documents, preserves names, distinguishes similar places, cites evidence and admits when information is absent. A model that sounds fluent but regularly returns the wrong source should not be trusted simply because its architecture is private.

Governance should enable choices

A good framework should make it easier to say yes to low-risk, beneficial uses and easier to say no to inappropriate ones. If every use requires the same approval path, people either avoid useful experimentation or find ways around the process. Neither outcome improves sovereignty.

Proportionality is therefore not a weakening of Māori data governance. It is a way of directing the strongest attention to the situations where authority, sensitivity and consequence genuinely demand it.

A practical recommendation

For each AI use, document four things in plain language: what information is involved, where every important component runs, who has authority over the use, and what can happen because of the output. Add a simple data-flow diagram and a statement about how evidence will be checked.

Then apply the appropriate governance. The result should be neither automatic approval nor automatic prohibition. It should be an informed decision about a real system.

Next in the series: From data sovereignty to AI capability.