Skip to main content

Why use local AI?

A cloud AI service can be convenient, but convenience is not the only question when the material concerns whenua, whānau, hapū, iwi, mātauranga or organisational information.

A better starting point is: where does the information go, what copies are made, what can be inferred from it, and who has authority over the use?

What local processing changes​

With a properly configured local setup, the main working parts can remain on your own computer:

  • the source PDFs and other files
  • the local language model
  • the embedding model used to find relevant passages
  • the document index or vector database
  • prompts and chat history
  • extracted tables and working outputs

That can reduce disclosure to an external AI provider and can make deletion, storage and access more tangible. It also lets you test the claim by disconnecting the network and checking that the work still functions.

This is a practical form of control. It can support rangatiratanga over the mahi because the organisation or person can decide where the working copies live and when they are removed.

What local processing does not change​

A model running on a laptop does not gain authority over Māori information. It can still misunderstand context, infer information that was not explicitly recorded, join fragments that should not be joined, or produce a confident interpretation that no appropriate person has approved.

Local processing does not answer questions such as:

  • Was the original information collected with authority?
  • Is the intended AI use within the original kaupapa?
  • Should this kōrero be indexed or summarised at all?
  • Who may see the extracted information?
  • Is the result appropriate to map?
  • Is an apparently public source actually culturally sensitive in context?
  • Who can confirm a local name, association or interpretation?

For those questions, use the wider Māori Data Sovereignty guidance and the Māori AI governance material.

Public does not mean unrestricted​

A useful distinction for historical and Treaty research is between being able to access a document and having good reason to reuse everything within it in any form.

Waitangi Tribunal reports, inquiry documents, Gazette notices, historical newspapers and other public sources can be excellent research material. They can also contain statements about whakapapa, wāhi tapu, urupā, customary use, disputed place names, old addresses or other information that becomes more revealing when extracted, joined and mapped.

Before turning public-source material into a new dataset, ask:

  1. What was the original source and kaupapa?
  2. Is the passage evidence, allegation, interpretation or settled finding?
  3. Does extracting it remove qualifications or surrounding context?
  4. Does adding coordinates create new sensitivity or apparent certainty?
  5. Is publication of the derived layer necessary for the mahi?
  6. What source reference should travel with the record?

The Waitangi Tribunal provides public reports and inquiry documents, but public access does not remove the need to preserve evidential context or think carefully about derived spatial outputs.

See Open data and research, Sensitive places and Provenance and source checking.

Cloud AI and local AI are not simple opposites​

Many desktop AI applications can connect to both local and cloud models. Some have optional web search, plugins, cloud model providers, update checks or telemetry. A desktop-looking interface therefore does not prove that every part of a workflow is local.

Before using material that needs stronger control, check the actual configuration:

  1. Identify the selected language model and where its model file is stored.
  2. Identify the embedding model and where it runs.
  3. Find the application data folder and document index.
  4. Check whether OpenAI, Anthropic, Google, OpenRouter or another cloud provider is configured.
  5. Check whether web search, external tools or third-party plugins are enabled.
  6. Disconnect Wi-Fi or the network and repeat a document question you already tested.
  7. Confirm that the answer and source retrieval still work.

If the workflow stops because a cloud model was selected, it was not a fully local workflow.

A useful Māori GIS example​

Suppose an iwi environmental team has a set of public planning reports and internal working notes about an awa. Local AI could help staff retrieve every passage discussing erosion, access or historical place references. The team could then review those passages and decide which information belongs in a GIS layer.

The AI can reduce the time spent finding text. It cannot decide what the iwi considers authoritative or appropriate to publish.

That distinction matters because GIS turns descriptions into spatial objects. A loose textual reference can become a precise-looking point or polygon very quickly. See Why GIS is different and Spatial inference risk.

Provenance still matters​

Keep the source visible as information moves through the workflow. At minimum retain the document title, page or section, the text that supports the extraction, the date of extraction, the model or method used, and the review status.

For ingoa wāhi, keep the spelling exactly as it appeared in the source before creating any normalised or official-name field. This avoids silently changing the historical evidence to match a modern spelling.

Delete means more than deleting a chat​

A local document-research application may hold several copies or derived representations of the same material. For example, AnythingLLM Desktop stores parsed document content, a local LanceDB vector database, cached embeddings, locally stored models and an SQLite application database under its desktop storage directory on Windows.

That is useful because the storage is inspectable. It also means a cleanup process should consider the workspace, parsed documents, vector index, direct uploads and exported results, not just the visible chat.

See Install local AI for the current Windows storage path and a practical local test.

Māori governance sources​

Te Mana Raraunga places rangatiratanga, whakapapa, whanaungatanga, kotahitanga, manaakitanga and kaitiakitanga at the centre of Māori Data Sovereignty. Te Kāhui Raraunga's Māori Data Governance Model provides a Māori-designed governance model, while its Māori AI Governance Framework extends that thinking into AI systems. The CARE Principles provide an international Indigenous governance framework centred on Collective Benefit, Authority to Control, Responsibility and Ethics.

Those principles apply whether the processing happens in a hyperscale cloud or on a laptop. Local infrastructure can strengthen practical control, but the kaupapa and authority remain with people.

Sources​

Last verified: 2 September 2026

Last updated on