AI for Māori GIS
Artificial intelligence is now appearing across mapping, GIS, imagery, document research and ordinary office software. Some of it is genuinely new. Some of it is an extension of machine-learning methods that GIS practitioners have used for years. The important thing is not to treat every product carrying an AI label as the same technology or the same risk.
For Māori GIS, AI can help search large collections of reports, find candidate ingoa wāhi, draft code, create metadata, structure field information, classify imagery, detect possible features and reduce repetitive technical work. It can also invent convincing details, infer sensitive locations, flatten uncertainty, expose information through external services or turn tentative evidence into a precise-looking point or polygon.
The site therefore treats AI as both a practical capability and an information-architecture question. The aim is to understand what the technology is doing, choose where it is useful, keep the source evidence visible and retain appropriate authority over the information and the resulting map.
New to AI?
Start with AI basics for Māori GIS. It explains AI, machine learning, deep learning, generative AI, GeoAI, foundation models, LLMs, training, inference, prompts, tokens, context windows, hallucination, embeddings, vector databases, semantic search, RAG, fine-tuning, multimodal models, computer vision, agents, tools, local and cloud AI, open-weight models, quantisation and the hardware used to run models.
The primer also explains two terms that can be particularly confusing in GIS. An AI vector is a mathematical list of numbers used for similarity, not a GIS point, line or polygon. A vector database used for RAG is therefore not a geodatabase or spatial database, even though both use the word vector.
Once those concepts make sense, the rest of the AI section becomes much easier to navigate.
Where AI fits into mapping
AI can enter a mapping workflow before, during or after conventional GIS analysis. Before mapping, it can search reports, extract candidate place names, compare planning documents, structure notes and help locate relevant evidence. During GIS work, it can draft QGIS expressions, Python or SQL, help diagnose errors and assist with data preparation. Machine learning and deep learning can classify spatial or tabular data and extract candidate features from aerial imagery, satellite imagery, video or point clouds.
The boundary is important. AI should not become the source of geographic truth simply because it can generate an answer. Coordinates, geometry, coordinate reference systems, statutory boundaries and authoritative spatial relationships still need real GIS sources and reproducible spatial methods. A language model can help find evidence for a place. It should not invent the coordinate from memory.
A useful principle across this site is: use GIS to establish the geography, use AI to assist with the information around the geography, and keep people with appropriate authority responsible for meaning and use.
Choose what you need
I want to understand the terminology
Use AI basics for Māori GIS. It is the general introduction and glossary for the rest of the site.
I want to understand training data and what happens to my prompts
Use AI training data, privacy and Māori GIS. It summarises how foundation models are trained, known Māori-related public corpora, evidence levels for likely training-data inclusion, current consumer and enterprise data-treatment differences across major AI providers, the Southland AI factory, and GIS-specific data-flow questions.
For the long-form research behind that reference, start with How AI actually learns and follow the six-part series through Māori training data, current AI services, US and Chinese model pipelines, Southland infrastructure and practical implications for Māori GIS.
I want to search PDFs and reports on my own computer
Go to Local AI and Māori GIS. It covers software choices, installation, offline testing, document search, RAG, prompts, place extraction and moving checked findings into GIS.
Before loading Māori or organisational information, read Before you use AI with Māori information.
I need governance, sovereignty and risk guidance
Use Artificial intelligence and GIS. It deals with processing location, retention, training and evaluation, derived information, inference, contracts, provenance and Māori Data Sovereignty.
For wider frameworks, see Further reading and frameworks.
I want practical GIS uses for AI
Use Practical AI applications for document extraction, tables, metadata, code assistance, field information and GIS preparation.
Use Find ingoa wāhi with local AI and Check the places for the complete source-checking workflow before a candidate name becomes spatial data.
I need to understand what can go wrong
Use What can go wrong for hallucination, weak retrieval, missing source context, OCR problems, false coordinates, cloud configuration mistakes, model limitations and other practical failure modes.
Use Inference risk when apparently ordinary data can reveal sensitive places or relationships once combined with location or other sources.
Training is not the same as use
A recurring source of confusion is the difference between training a model and using a trained model. Training changes the model. Inference is the act of asking an already-trained model to perform a task. A RAG system usually retrieves passages from your documents and supplies them to the model during inference rather than training the model on the document collection.
That distinction does not make every cloud service appropriate. A provider may still process, retain, log or review submitted material according to its product terms and configuration. It does mean the technical question should be specific: what exactly happens to this information in this architecture?
Local inference can materially change that answer because the model and document processing can run on hardware controlled by the organisation. It does not decide whether the underlying information was appropriate to index, whether a model's original training was culturally appropriate or whether an output has authority.
Training-data history matters too
Public Māori material has already entered at least some web-scale AI corpora. The mC4 Māori-language subset, derived from Common Crawl, contains identifiable Te Ara, Māori media, cultural and iwi-related pages among roughly 101,000 records automatically classified as Māori. That number includes classification errors and should not be treated as 101,000 clean Māori documents, but it provides direct evidence of the public-web-to-training-data pathway.
Māori-related information is also much larger than the te reo Māori portion of a corpus. Waitangi Tribunal reports, settlement documents, environmental plans, academic research and other public material may contain detailed Māori information while being classified as English. The AI training data and privacy reference separates demonstrated, likely, plausible and unknown inclusion so that the site does not overclaim what a particular model has seen.
Historic foundation-model training and present-day data use need different responses. An organisation may not be able to remove every influence of an old public webpage from a model that has already been trained, but it can still control whether a new internal report is uploaded to a consumer service, placed under enterprise terms or searched entirely with local AI.
Five checks before using AI
Before putting information into an AI workflow, identify what the system will actually receive, where each component runs, who has authority or legitimate interests in the information, what the model or agent could infer or create, and how the output will be checked before it becomes GIS data, a map or a decision input.
These checks should be proportional to the use. Asking a local model to explain a QGIS error using synthetic data is not the same as building a system that can search restricted historical material or make consequential recommendations about people. Architecture, information sensitivity and consequence all matter.
AI-generated maps
Generative image tools can produce attractive map-like graphics without geographic integrity. Treat those outputs as illustrations unless they are built from real geometry and a reproducible spatial process.
For geographic or decision-support maps, retain the real geometry, coordinate reference system, source datasets, method, legend and provenance. A screenshot or generated image does not carry the same geographic information as the underlying GIS data, even when it looks convincing.
AI is not cultural authority
A model can retrieve passages, compare text, draft code, classify imagery and identify candidate patterns. It does not acquire authority over whakapapa, tikanga, local history, ingoa wāhi or relationships because it can generate fluent text about them or because it runs on Māori-controlled hardware.
Keep generated interpretation separate from source evidence. Where meaning depends on people and relationships, the relevant people remain the source of authority. AI can help reach the evidence faster without replacing those relationships.
Current Māori-led guidance
Useful starting points beyond MāoriGIS.nz include Te Kāhui Raraunga's Māori AI Governance Framework, its Māori AI Governance FAQ and Māori Data Governance Model, along with Te Mana Raraunga and the Global Indigenous Data Alliance CARE Principles.
AI products, terms and model behaviour change quickly. Check current provider information when choosing a service and revisit the decision when the tool, data, purpose, audience or level of automation changes.
For the longer practical discussion about architectures, local AI and capability, the earlier MāoriGIS.nz blog series begins with Where does your data actually go? Six ways to use AI. The newer research series on training data and current provider behaviour begins with How AI actually learns.
Last reviewed: 4 September 2026