AI basics for Māori GIS

Artificial intelligence is not one tool and it is not one kind of computer system. The term covers a broad family of methods that can recognise patterns, classify information, make predictions, generate text or images, extract features and assist with decisions. Some AI has been used in GIS for years. The newer wave of generative AI adds language models, multimodal models, document search and agents that can interact with software and data.

For Māori GIS, it helps to understand the parts before deciding whether a particular use is useful, safe or appropriate. A model classifying vegetation from aerial imagery is doing something quite different from a chatbot summarising a Treaty report. A local language model searching documents on an iwi-controlled computer is different again from uploading the same material to a public cloud service.

This page explains the main concepts in plain English and connects each one to mapping. It is not a governance framework and it does not decide who has authority over Māori information. The purpose is to make the technology understandable enough that governance, technical design and GIS practice can refer to the same actual system.

The big picture

A useful mental model is to treat artificial intelligence as the broad umbrella. Machine learning sits within that broad field and learns patterns from data rather than relying only on rules written by a programmer. Deep learning is a type of machine learning built around large neural networks. Modern generative AI usually uses deep learning to create new text, images, audio, code or other content.

GeoAI, or geospatial artificial intelligence, is not a separate species of AI. It is the application of AI together with geospatial data, spatial analysis and geographic science. Esri describes GeoAI as the fusion of AI with geospatial data, science and technology. In practice that can mean detecting buildings in imagery, classifying land cover, forecasting change, extracting locations from text or helping a GIS practitioner write code.

Traditional GIS analysis does not become AI simply because a computer performs it. A buffer, overlay, spatial join, coordinate transformation or viewshed follows a defined spatial method. Machine learning becomes useful when the task involves learning patterns, ranking possibilities, prediction, classification or extracting information from complex data. Generative AI becomes useful when the task involves language, images, code, documents or interactive assistance.

AI, machine learning and deep learning

Machine learning learns relationships from examples. A model might learn from historical environmental measurements and then estimate erosion risk in other places. A classification model might learn which combinations of attributes are associated with a particular land-cover class. The model does not need a programmer to write an explicit rule for every possible case, but it still depends heavily on the quality and representativeness of its training data.

Deep learning uses much larger neural networks and is especially useful for complex data such as imagery, speech and text. In GIS, deep learning is commonly used for object detection, image classification and segmentation. A model might identify buildings, roads, vegetation, slips or other candidate features in aerial or satellite imagery and return a raster or vector layer for review.

The important GIS discipline remains the same: the model output is a derived dataset. A precise polygon created by a deep-learning model is not automatically a verified building, archaeological feature or culturally significant place. Keep the source imagery, method, model information and review status with the result.

Generative AI

Generative AI creates new content from patterns learned during training. That content can include text, code, images, audio, video and structured data. ChatGPT, Claude, Gemini and Microsoft Copilot are familiar examples of applications built around generative models, but the underlying model and the application are not the same thing.

For GIS, generative AI is useful for jobs such as explaining a QGIS error, drafting a Python script, writing a SQL query, summarising reports, extracting candidate place names, creating draft metadata, comparing planning documents or turning reviewed text into a proposed table structure. It can also describe images and help interpret technical documentation.

Generative AI is much less reliable as an authority for geography. Asking a language model to invent coordinates, boundaries or the history of a place from memory can produce very convincing errors. Use authoritative spatial sources and GIS methods to establish the geography, then use AI to assist with the information around it.

Models, applications and services

A model is the learned mathematical system that performs the AI task. An application is the software interface people use. A service is the infrastructure through which the model or application is delivered.

Those three layers are easy to blur. AnythingLLM, LM Studio, ChatGPT and Copilot are applications or services. A particular GPT, Gemini, Claude, Llama, Qwen or other model sits underneath an application. The same application may sometimes connect to several different models, and the same model may be available through several different services.

This distinction matters for Māori data sovereignty because the application name does not tell you where processing happens. A desktop interface may be using a local model, a remote cloud model or a mixture of both. The useful questions are where the model runs, where source data and indexes are stored, what leaves the organisation and what is retained.

Foundation models and LLMs

A foundation model is trained broadly enough that it can be adapted or prompted for many different tasks. Large language models, usually shortened to LLMs, are foundation models designed primarily to work with language. They learn statistical relationships between pieces of text and generate responses one token at a time.

An LLM does not contain a normal database of facts that it looks up each time it answers. Its training has changed a very large set of numerical parameters so the model can generate likely and useful continuations from the context it receives. This is one reason it can answer questions fluently and still be wrong.

Modern reasoning models are language models that have been trained and tuned to perform stronger multi-step problem solving. They can be very capable with code, planning and analysis, but they are still models generating outputs from learned patterns and supplied context. More reasoning does not turn them into an authoritative source for a place name, boundary or cultural interpretation.

Training and inference

Training is the process used to create or change a model. Large foundation models are initially trained on very large datasets. They can later receive additional training, fine-tuning or other forms of post-training to improve behaviour or specialise them.

Inference is what happens when a trained model is used. You provide a prompt, an image, a document extract or some other input and the model produces an output. When a local model answers a question on your computer, that is inference.

Keeping training and inference separate is important. Giving a document to a model for inference does not automatically mean that the document becomes part of the model's future training. Whether a cloud provider retains or uses submitted material for model improvement depends on the product, account type, configuration and current terms. Running inference locally can remove the need to transmit working documents to an external model provider, but it does not answer questions about the model's original training data.

Prompts, tokens and context

A prompt is the information supplied to a generative model for a particular request. It can include the user's question, instructions, retrieved source passages, examples, images and system rules. Prompting changes what the model sees for that request. It normally does not change the trained model itself.

Language models process text as tokens rather than as whole documents or whole words. A token is a small unit of text. The amount of information a model can consider at once is limited by its context window. A long context window allows more material to be supplied in one request, but more context does not guarantee better reasoning or better evidence retrieval.

For GIS research this matters when people expect a model to understand a folder containing thousands of pages. The system usually needs a search or retrieval layer to identify the relevant passages rather than sending the entire archive to the model for every question.

Hallucination

Hallucination is the common term for an AI output that sounds plausible but is unsupported or false. The model is not necessarily attempting to deceive. It is producing a response that fits the learned and supplied patterns even when the evidence is missing or uncertain.

GIS provides particularly dangerous examples because false information can look precise. A model can produce a realistic latitude and longitude, a confident place-name explanation or a plausible-looking boundary that is simply wrong. The normal protection is not more confidence in the model. It is provenance: return to the source, check the geometry and preserve the evidence used to create the GIS feature.

Embeddings

An embedding is a numerical representation that places text, images or other information into a mathematical space where similar items can be found near one another. Embeddings make semantic search possible because a system can find material that is related in meaning even when the exact words are different.

For example, a search for river restoration may retrieve passages discussing riparian planting, bank erosion or freshwater rehabilitation even when the phrase river restoration does not appear exactly. Multilingual embedding models can also help find related material across languages, although their performance with te reo Māori, historical spellings and specialist terminology still needs testing.

There is an important GIS terminology trap here. An AI embedding vector is not the same thing as GIS vector data. In GIS, vector data means points, lines and polygons with geometry. In AI, a vector is simply a list of numbers representing features or meaning. A vector database stores those numerical representations for similarity search. A geodatabase or spatial database stores geographic geometry and attributes. They solve different problems.

Vector databases and semantic search

A vector database stores embeddings and helps retrieve the items that are mathematically most similar to a query. It is commonly used in document-search systems. The original text is usually split into smaller chunks, each chunk receives an embedding and the embeddings are indexed for retrieval.

When someone asks a question, the question is also converted into an embedding. The system finds the closest document chunks and returns them as possible evidence. The language model then works with those retrieved passages.

This is different from ordinary keyword search. Keyword search is often stronger when you need an exact historical spelling, parcel identifier or uncommon name. Semantic search is stronger when the concept may be expressed in several different ways. Good research systems often need both.

RAG

RAG means retrieval-augmented generation. The phrase sounds more complicated than the idea. The system finds relevant information first and then gives that information to the language model so the answer can be grounded in the retrieved evidence.

A simple document RAG workflow looks like this:

Documents
   ↓
extract and split text
   ↓
create embeddings
   ↓
search index / vector database
   ↓
retrieve relevant passages
   ↓
add passages to the prompt
   ↓
language model
   ↓
candidate answer with sources

RAG normally does not retrain the language model on your documents. It supplies selected information at inference time. That distinction is particularly useful for Māori organisations because documents can remain in a controlled collection while the retrieval system decides what evidence the model receives for each question.

RAG can still fail. The wrong page may be retrieved, OCR may have damaged the text, the embedding model may be weak for te reo Māori, or the language model may ignore part of the evidence. A RAG answer should therefore remain linked to the source passage rather than being treated as a new authoritative record.

Fine-tuning is different from RAG

Fine-tuning changes the model itself by additional training. It can help adjust behaviour, style or performance for particular classes of task. It is not normally the best way to make a language model remember a changing collection of organisational documents.

RAG is usually a better fit when the aim is to search reports, policies, historical documents or metadata and keep answers tied to current sources. Fine-tuning is more appropriate when an organisation has a well-defined model behaviour it wants to teach and has suitable training and evaluation data.

For Māori information, fine-tuning also raises a different governance question because the information may become part of the model weights or training artefacts. That is a different architecture from retrieving approved passages at question time.

Multimodal AI

A multimodal model can work with more than one kind of input or output, such as text and images, or text, images and audio. This is useful in GIS because geospatial work rarely consists of text alone. It may involve aerial imagery, scanned historical maps, photographs, diagrams, tables and written reports.

A multimodal model can describe an aerial image, extract text from a scanned map or compare two photographs. That does not mean it understands the map as a GIS dataset. A screenshot of a map is an image. It does not carry the coordinate reference system, attribute table, topology or exact geometry available in the original GIS layer.

When spatial accuracy matters, keep the real GIS data and use image-capable AI as an assistant rather than converting the entire workflow into pictures.

Computer vision and imagery AI

Computer vision is AI applied to images and video. In geospatial work it can classify imagery, detect objects, segment areas and identify change. Modern GeoAI tools can extract buildings, roads, vegetation, water, erosion features and other candidate spatial information from imagery, video and point clouds.

These workflows are often more structured than generative chat. A model processes pixels and returns classes, bounding boxes, masks, probabilities or other derived information. Those outputs can then become GIS layers.

For taiao or historical work, this can be powerful, but the same provenance rule applies. Keep the source imagery, model, date, parameters and human validation. A feature discovered by a model is a candidate until the relevant evidence and people support treating it as something more.

Agents, tools and connectors

An AI assistant answers. An AI agent can also take actions. Modern agent systems can call tools, search databases, run code, read files, use web services and repeat steps until a task is complete. A tool is simply a capability the model is allowed to invoke. A connector gives the AI access to another system or data source.

In GIS, an agent could potentially search documents, query a spatial database, call a geocoding service, run Python, create a table and launch a geoprocessing step. That can remove repetitive work, but it also expands the data boundary and the consequences of a mistake. A model that only drafts a QGIS expression cannot change the database. An agent with credentials and write access potentially can.

The more an AI system can act, the more important permissions, logging, approval points and human review become. Agentic AI should not be treated as just a more powerful chatbot.

Cloud, enterprise, local and hybrid AI

Cloud AI runs the model on infrastructure operated by another organisation. Public chatbots are the most visible example. Enterprise cloud AI may add contractual protections, identity controls, audit functions and different rules about how organisational data are handled.

Local AI runs the model on hardware controlled by the user or organisation. A fully local document system can also keep the document parser, embeddings, vector database and chat history local. This can materially reduce disclosure to external AI providers.

Hybrid systems combine local and cloud components. A common example keeps the full document collection locally but sends selected retrieved passages to a cloud model. That may be a sensible architecture, but it is not fully local. The retrieved passages still cross the boundary.

For Māori GIS, draw the data flow rather than relying on the label. Show where the GeoPackage, geodatabase, imagery, field notes, PDFs, embedding model, index and language model actually sit. A final public map can be harmless while the source layers beneath it are highly sensitive.

Open-weight, open source and proprietary models

A proprietary model is controlled and hosted under the terms of its provider. An open-weight model makes its trained numerical weights available so it can be downloaded and run on infrastructure you control. Open-weight does not automatically mean that all training data, training code and development decisions are open.

The phrase open source is used inconsistently in AI, so it is useful to read the actual licence and model documentation instead of relying on the label. OpenAI, for example, describes its gpt-oss models as open-weight models that can run on infrastructure controlled by the user. Other model families use different licences and release approaches.

For sovereignty, open-weight models can increase infrastructure and operational choice because they can be run locally or on Māori-controlled servers. They do not automatically solve questions about training provenance, bias, cultural authority or software security.

Parameters, quantisation, RAM and GPUs

A model contains a very large number of learned parameters. In broad terms, larger models often have more capacity, but model quality also depends on training, architecture and the task. Parameter count alone is not a reliable ranking system.

Quantisation stores model numbers at lower precision so the model uses less memory and can run on smaller hardware. This is one reason useful language models can now run on ordinary workstations and some laptops.

RAM is general system memory. VRAM is memory on a graphics processing unit, or GPU. GPUs can accelerate AI workloads because they are good at the large parallel calculations used by neural networks. A model can often run on a CPU without a dedicated GPU, but generation may be slower. Context length, model size and other applications also consume memory, so the model file size is not the same as the total RAM required.

For a first Māori GIS document-search pilot, the objective is not to run the largest model available. A smaller model that retrieves the right evidence, preserves source references and responds promptly may be far more useful than a much larger model that is slow and difficult to operate.

Where AI fits into mapping

AI can enter a GIS workflow at several different points. Before mapping, it can search reports, extract candidate place names, identify relevant planning provisions, draft metadata and structure field notes. During analysis, machine learning can classify or predict from spatial and tabular data, while computer vision can extract candidate features from imagery. During technical GIS work, a language model can explain errors, draft expressions, help with SQL or Python and suggest processing steps. During communication, AI can help draft plain-language descriptions, accessibility text or summaries of a map's evidence.

The strongest workflows keep the roles separate. GIS remains responsible for coordinates, geometry, coordinate reference systems, spatial analysis and reproducible geographic outputs. AI helps with search, extraction, classification, coding, language and pattern recognition. People remain responsible for checking evidence, deciding meaning and determining whether information should be mapped or published.

A useful document-to-map workflow is:

source documents
   ↓
AI search and candidate extraction
   ↓
human source check
   ↓
geographic source check
   ↓
structured GIS fields
   ↓
verified geometry
   ↓
map

For imagery, the sequence is different:

source imagery
   ↓
AI classification or feature detection
   ↓
candidate raster / points / polygons
   ↓
human or field validation
   ↓
reviewed GIS layer
   ↓
analysis or map

What AI should not replace

AI should not replace a coordinate source simply because it can produce a latitude and longitude. It should not replace authoritative cadastral or statutory boundaries. It should not silently convert an uncertain historical description into a precise point. It should not overwrite source spelling with a modern form without preserving the original. It should not replace the people who hold authority over tikanga, whakapapa, local history or the appropriate use of sensitive information.

This is not an argument for avoiding AI. It is a division of labour. Use AI where it reduces repetitive effort, expands search capability or helps with complex technical tasks. Use GIS where geographic integrity matters. Use people and appropriate authority where meaning, consequence and relationships matter.

Māori data sovereignty and AI architecture

Māori data sovereignty remains relevant across every architecture, but the questions change depending on the system. It is useful to separate authority over the information from control of the infrastructure and from authority over interpretation.

A fully local model can materially strengthen control over where organisational data are processed, but local hosting does not decide whether the information should have been indexed or analysed. An enterprise cloud service may provide strong contractual and security controls while still processing information on external infrastructure. A model may be open-weight while its original training provenance remains incomplete. A technically accurate output may still lack cultural authority.

For practical work, start by following the data. Identify the source information, storage, model, embeddings, index, tools, logs and outputs. Then ask who controls each part, what crosses organisational boundaries, what new information can be inferred and who remains accountable for the result.

See Before you use AI with Māori information and Artificial intelligence and GIS for the detailed governance guidance.

A short glossary

Term	Plain-English meaning
AI	The broad family of computer methods that perform tasks involving learning, prediction, perception, generation or decision support.
Machine learning	AI that learns patterns from data rather than relying only on fixed rules.
Deep learning	Machine learning using large neural networks, especially useful for imagery, language and other complex data.
Generative AI	AI that creates new text, code, images, audio or other content.
GeoAI	AI used with geospatial data, GIS and spatial science.
Model	The trained mathematical system that performs the AI task.
Foundation model	A broadly trained model that can be adapted or prompted for many downstream tasks.
LLM	A large language model designed to work with text and language.
Training	The process that creates or changes model parameters using data.
Inference	Using an already-trained model to produce an output from a new input.
Prompt	The instructions and context supplied to a generative model for a request.
Token	A small unit of text processed by a language model.
Context window	The finite amount of prompt, retrieved evidence and conversation a model can consider at once.
Hallucination	A plausible-looking AI output that is unsupported or false.
Embedding	A numerical representation used to measure similarity between text, images or other information.
Vector database	A database designed to store and search embeddings. It is not the same as GIS vector geometry.
Semantic search	Search based on similarity of meaning rather than exact words alone.
RAG	Retrieval-augmented generation: find relevant evidence first, then give it to the model for the answer.
Fine-tuning	Additional training that changes a model for a particular behaviour or task.
Multimodal	A model that can work with more than one type of information, such as text and images.
Computer vision	AI for recognising, classifying or extracting information from images and video.
Agent	An AI system that can use tools and take actions as well as generate an answer.
Tool	A function or service an AI system is allowed to call, such as search, code execution or a GIS service.
Connector	A link between an AI system and another application, account or data source.
Local AI	AI inference running on hardware controlled by the user or organisation.
Cloud AI	AI inference running on infrastructure operated by an external provider.
Open-weight model	A model whose trained weights can be downloaded and run on infrastructure you control.
Quantisation	Reducing numerical precision so a model uses less memory and is easier to run locally.

Where to go next

If the terminology now makes sense, return to AI for Māori GIS for the site-wide route into practical use, governance and risk guidance. For hands-on work, continue to Local AI and Māori GIS. For examples tied directly to mapping, see Practical AI applications.

For the broader architecture discussion, the MāoriGIS.nz blog series begins with Where does your data actually go? Six ways to use AI.

Useful external references include the NIST AI glossary, Esri's GeoAI overview, Google's RAG overview and current model/provider documentation for any system you are considering.

Last reviewed: 4 September 2026