Skip to main content

Before you use AI with Māori information

Local AI can be a useful way to reduce disclosure of working documents to external AI providers. It is one technical control, not a complete governance answer.

Before loading a collection of reports, maps, photographs, whakapapa-linked records, ingoa wāhi or organisational information into any AI system, consider the kaupapa, the source, the people with authority or interests in the information, the system being used and the new information the AI may create.

Local processing can support greater control. It does not by itself make an AI use sovereign, authorised, culturally appropriate or low risk.

Publicly available is not the same as unrestricted​

Public reports, historical records, inquiry evidence, newspapers, maps and government datasets can be valuable sources. Public access does not necessarily mean that every item of kōrero is appropriate to extract, combine, interpret, map or republish without further thought.

A public source can contain whakapapa, historical statements, wāhi tapu, urupā, customary relationships, disputed evidence, approximate locations or material provided for a particular inquiry. AI and GIS can make that information easier to aggregate and can create new spatial relationships that were not obvious in the original documents.

Keep the source context. Ask whether a derived dataset or map changes the accessibility, sensitivity or apparent certainty of the original information.

Check your own organisation first​

If you are working for an iwi organisation, trust, Māori land organisation, government agency, council, research institution, company or other employer, check the rules that apply in that environment before installing software or loading organisational information.

Depending on the organisation and the information, that may include:

  • Māori data governance or tikanga guidance
  • information security requirements
  • privacy requirements and Privacy Impact Assessment processes
  • legal and records-management requirements
  • software approval and endpoint-management rules
  • procurement and supplier assurance
  • cloud and data-location policies
  • research ethics or data-sharing agreements
  • contractual obligations to whānau, hapū, iwi, partners or data providers

A local desktop application can still create document copies, embeddings, indexes, logs and derived outputs. It can also include optional cloud providers, plugins, telemetry or web tools. Know the actual configuration rather than relying on the word local in a product description.

Check the model as well as the application​

Where possible, read the model card, licence and provider documentation for the model itself. Consider who produced it, what the provider discloses about training and evaluation data, known limitations, supported languages and whether there are restrictions on use.

A model being downloadable or open-weight does not mean its training data are fully documented, culturally representative or suitable for interpreting Māori information. A model may perform well at retrieval and summarisation while still being weak with te reo Māori, historical spellings, tikanga, whakapapa or local context.

Keep two questions separate:

  • what data were used to create or train the model?
  • what happens to the documents and prompts you give the model during your own use?

Running inference locally can improve control over the second question. It does not resolve the first.

Start with Māori-led frameworks​

MāoriGIS.nz provides practical GIS guidance, but readers should go beyond this site and consider Māori-led frameworks directly.

Do not treat these links as a one-off reading list. The field is developing quickly. Revisit the source organisations, look for newer guidance and consider whether it changes the approach to your mahi.

Other useful Aotearoa guidance​

Different sources answer different parts of the problem. Depending on your role, also consider:

These frameworks overlap, but they are not interchangeable. A privacy assessment does not answer Māori Data Sovereignty questions. Keeping a model offline does not answer tikanga or authority questions. A Māori governance framework does not remove the need for cyber security, procurement or legal review where those requirements apply.

Questions before starting​

For ordinary document research, ask at least:

  1. What is the kaupapa and expected benefit?
  2. What information will be loaded, indexed or generated?
  3. Who has authority, rights or legitimate interests in that information?
  4. Is the material public, internal, confidential, culturally sensitive or mixed?
  5. Does public availability hide a more complicated cultural or evidential context?
  6. Where will source files, parsed text, embeddings, prompts and outputs be stored?
  7. Is any part of the workflow using an external model, embedding API, plugin, connector or web service?
  8. What does the software retain, log, transmit or use for product improvement?
  9. Who produced the model, what is known about its training and evaluation data, and what limitations are documented?
  10. What new information might the model infer, reconstruct or reveal by combining sources?
  11. How will AI findings be checked against the original source?
  12. Who remains accountable for accepting, rejecting or acting on an AI-generated finding?
  13. Who decides whether a candidate ingoa wāhi, relationship or interpretation is accepted?
  14. Is mapping the result appropriate, and should the map be private, generalised or public?
  15. What organisational approvals or assessments apply?
  16. What will be deleted, retained, archived or handed over when the mahi ends?

The answers may lead to a fully local workflow, an approved enterprise service, a non-AI search process, a smaller source collection, tighter access controls, consultation with relevant people, or a decision not to digitise or map particular kōrero at all.

Keep researching as the mahi develops​

Do not assume the research question is settled once the software is installed. Reconsider the sources and controls when the collection changes, a new dataset is joined, a model or application is upgraded, information becomes more sensitive through aggregation, or the output moves from private research to publication.

For a significant kaupapa, look beyond generic AI guidance. Seek the policies, tikanga, data-sharing arrangements, mātauranga guidance, iwi or hapū expectations, legal requirements and specialist advice that apply to that particular context. If new guidance changes the risk or the appropriate use, change the workflow.

Continue on MāoriGIS.nz​

Last verified: 2 September 2026

Last updated on