The map is still the easy part: managing AI in Māori GIS
A few years ago the AI part of a GIS workflow was simple enough. Open ChatGPT, type a question, see what comes back. Sometimes it was useful. Sometimes it confidently invented a Python function. At least the administrative overhead was low.
That has changed. A Māori GIS practitioner wanting help with an ordinary piece of work can now face a small set of decisions before the work starts. ChatGPT, Claude, Gemini, Copilot or something more specialised? A fast model or the one that thinks for longer? Ordinary chat, a project, a research mode, a coding agent or a persistent workspace? Is the task worth using a limited premium allowance? Does the system need three source documents or the whole project folder? Can it use the web? Can it run code? Will it remember anything later? Most importantly for Māori GIS, what information are we about to give it, and should that information leave the environment it is already in?
This is starting to look like a professional skill in its own right. Knowing AI no longer means being good at typing questions into one chatbot. It increasingly means knowing how to select and manage several different kinds of machine assistance, give them the right amount of context, use expensive capability where it matters, keep data under appropriate control, and check the result before it quietly becomes part of the GIS.
There is no especially convincing name for this skill yet. “Prompt engineering” is too narrow because the prompt is only one part of the job. “AI orchestration” sounds expensive. “Model literacy” is better, although the model itself is only one item among workspaces, agents, context, connectors, local tools, usage limits and organisational rules. “AI operational literacy” is probably close enough until somebody invents a better term and changes it twice before the training material is printed.
For Māori GIS the skill has an extra dimension. A public LINZ layer, an openly published Waitangi Tribunal report, a synthetic parcel table and an unpublished set of culturally sensitive locations are not just four files of different sizes. They carry different authority, context and obligations. A model can be technically excellent and still be the wrong place to put the data. That decision belongs before model selection, not after it.
How much AI does the mahi need?
The easiest habit to acquire with AI is also one of the least useful: selecting the strongest model every time. If the most capable model is available, using anything less can feel like deliberately asking for a worse answer. In practice, the task matters more than the ranking.
The current products make this distinction increasingly explicit. As at August 2026, ChatGPT exposes different GPT-5.6 models and reasoning levels across Chat, Work and Codex. Gemini has faster and more capable model classes and lets eligible users select different levels of thinking. Claude usage changes with model choice, effort, conversation length, files and tools. GitHub Copilot now meters many chat and agent activities through AI Credits, while ordinary code completions on paid plans are treated differently. These are not comparable units and they should not be collapsed into one imaginary “AI credit” scale. A context limit is not a subscription limit. An API token price is not a weekly consumer allowance. A research run is not necessarily metered like a chat. Anyone paying for several services soon discovers that AI billing has achieved the rare feat of making GIS licensing look briefly familiar. Current details are worth checking in the provider documentation rather than memorising because they change regularly: OpenAI, Gemini, Claude and GitHub Copilot.
The practical response is not to become an accountant of reasoning tokens. It is to recognise different classes of work. Asking for twenty ideas for a beginner QGIS exercise is cheap reconnaissance. Reformatting a public data dictionary, drafting field aliases or turning a list of requirements into a test table does not need the largest model available. If the job is a difficult PyQGIS or ArcPy failure involving an enterprise geodatabase, software-version differences and several modules, stronger coding capability and current technical documentation become much more valuable. Researching a current law, policy or government programme is different again because currency and evidence matter more than how elegantly the model can speculate.
Māori GIS supplies plenty of examples. AI can help search public Waitangi Tribunal material for candidate ingoa wāhi, but the difficult part is not producing a list of names. It is retaining the source, separating the wording in the report from later interpretation, checking whether a national Gazetteer match is actually the right place, and recognising that a locally held or historical name does not become false because it is absent from a national database. A strong research model may help locate material quickly. It does not acquire local authority by reading it quickly.
The same applies to coding. An AI can write Python to compare Māori Land Spatial Dataset records with LINZ parcel information, create a QGIS expression, generate SQL for a PostGIS table, or explain why a spatial join is duplicating rows. These are good uses because the output is inspectable and testable. An extremely capable model can still be defeated by an unexpectedly named field, a null value, an outdated API method or the simple fact that the real schema differs from the schema described in the prompt. GIS software has always had the ability to turn one small incorrect assumption into an impressive quantity of output. AI has joined that tradition.
There is also no need to give a cloud model real information when a synthetic example will do. If the problem is designing a schema for recording restoration monitoring, give the AI ten invented records with the same field types. If the job is developing a script for a sensitive layer, give it the schema and fake values, review the code, then run the tested script locally against the real data. If a model needs to know the exact wāhi tapu coordinates in order to explain how a CSV import works, the workflow has gone wrong somewhere.
This is one reason local AI matters for Māori GIS. Running a model and document search on a computer you control can reduce disclosure to external services and can be particularly useful for document-heavy research. It does not make the AI culturally authoritative and it does not settle questions of tikanga, provenance or permission. Local processing changes an important technical fact, where the information is processed, but it does not turn every possible use into an appropriate one. The useful principle is still to provide the minimum information needed for the job.
Experienced users increasingly end up allocating AI much as they allocate other technical resources. Use the lightweight model for the mechanical work. Spend stronger reasoning on the difficult decision. Use the coding environment when the problem is code. Use research mode when current external evidence is genuinely required. Keep sensitive data local where that is the appropriate boundary. Escalate when the task earns it rather than starting everything at maximum power.
There is some humour in watching a profession that can transform coordinates between three reference systems without concern spend ten minutes deciding whether a paragraph about metadata deserves High reasoning. It is nevertheless a rational problem. Premium model allowances, agent credits and research limits are resources. The mistake is not thinking about them at all. The other mistake is spending more time optimising the model choice than the original task would have taken.
Context has consequences
The second skill is deciding what the AI should know. This is often described as context management, which makes it sound like a technical problem involving token counts. It is partly that, but for Māori GIS it is also an information-management and sovereignty problem.
Modern systems can accept remarkable quantities of material. That is useful when the task genuinely needs a long report, several code files or a collection of source documents. It also creates a very tempting button marked, in effect, “upload the drive”.
More context is not automatically better context. GIS practitioners already know this from data. Adding another twenty layers does not necessarily improve the analysis. Sometimes it just makes the contents panel longer. Context has the same problems as any other information collection: duplication, stale versions, conflicting instructions, weak provenance and material that was never relevant to the question in the first place.
A project containing three authoritative documents may be far more useful than one containing 300 files accumulated over six years. The large collection may include the current requirements, the old requirements, the earlier draft that everybody rejected, a spreadsheet copied from somewhere nobody can remember and the original meeting note that misspelled the field name. A large context window is perfectly capable of reading all of them. That is not the same as knowing which one governs the work.
For Māori information the consequences go further. Combining datasets can reveal relationships that were not obvious in the individual files. Coordinates, land records, imagery, names, historical references and public information can produce inferences that become more sensitive than any single source. Removing the NAME field does not necessarily anonymise a spatial dataset when the geometry identifies the place. Inference risk is already a data-sovereignty issue before AI becomes involved; AI simply makes combination and interpretation much easier.
This is where the habit of treating a chat history as a project repository becomes particularly poor. A conversation is useful working space. It is not the authoritative home for a dataset, a decision, a source register or a final methodology. If the accepted schema matters, keep the accepted schema. If a report is authoritative, keep the report. If a place-name match has been reviewed, record the review status and evidence in the dataset rather than trusting that the AI will remember what happened forty messages ago.
Persistent projects and workspaces are useful precisely because they can keep related files and instructions together. They still need curation. Old instructions should be removed or clearly superseded. Source documents should be identifiable. A project that contains every version of everything eventually develops archaeology. Somebody later has to work out why the AI still insists on calling a layer Sites_Final_2 when the organisation stopped using that name in April.
For iwi, hapū and whānau projects there is another question that should be asked before “how much context can this model hold?”: who has authority to place this material in this system? Possession of a file does not automatically confer permission to send it to another service. The existing Māori GIS material on this site makes that distinction deliberately. Māori Data Sovereignty for GIS and AI and Māori GIS both treat data movement, authority and secondary use as design questions rather than an appendix to the technical work.
That can lead to hybrid workflows rather than a single rule of “cloud bad, local good”. A public source can be researched with a web-connected system. A cloud coding assistant can work against synthetic data. Sensitive documents can be searched with a local model. A human can review the candidate findings and then write only the approved result into the GIS. The best AI architecture may therefore be several small boundaries rather than one enormous intelligent workspace containing everything anyone might conceivably need.
This is less glamorous than buying the most capable model. It is also much closer to good GIS practice.
Good prompts are mostly good decisions
Prompting still matters, but its importance is sometimes exaggerated because it is the visible part of AI use. The easiest thing to teach is a clever phrase to type into a box. The harder skill is deciding what the box actually needs to know.
“Act as an expert in Māori GIS” is not useless, but it is not much of a specification. The model does not acquire relationships, local knowledge or authority because the role instruction was written confidently. Better prompts define the work: the objective, relevant context, source material, constraints, intended audience, required output and how the result must be checked.
Consider a task that fits the local AI workflow: finding candidate ingoa wāhi in public historical reports.
A weak prompt would be:
Find all the Māori place names in these reports and map them.
That sounds efficient and contains several hidden decisions. Does “place name” include variants and historical spellings? Is the model allowed to modernise a name? Should it infer a location from context? What happens when several Gazetteer features have similar names? Is absence from the Gazetteer treated as failure? Does “map them” mean inventing coordinates for anything unresolved?
A more useful version would be:
Search only the attached public reports for candidate ingoa wāhi. Preserve each name as written in the source and record the report title and page. Summarise the surrounding geographic context without inventing a coordinate or modern equivalent. Keep duplicate spellings and variants visible. Return a table with
source_name,source_page,source_name_as_written,context,possible_modern_match,match_statusandnotes. Usematch_statusvalues ofunreviewed,possible,ambiguousorno_match. Do not treat absence from the New Zealand Gazetteer as evidence that a name is wrong. Do not assign coordinates until a geographic source has been checked separately.
The improvement is not the length. It is that several important professional decisions have been made before the AI starts. The output is designed for review rather than for the appearance of completion.
This is especially useful where an AI system tends to fill gaps. Generative models are built to continue patterns. GIS research often needs the opposite behaviour: stop here, mark this unresolved, retain the ambiguity, and do not turn a plausible association into geometry. A good prompt tells the system where not to be helpful.
One of the more useful techniques is to ask AI to design the prompt before running the task. Explain the outcome, provide the known constraints, then ask: “Do not do the task yet. Write the prompt I should use. Identify missing information, possible failure modes, evidence requirements, privacy issues, stopping conditions and acceptance tests.”
This is not as absurd as it first appears. Prompt design is itself a form of requirements analysis, and models can be good at spotting omitted conditions. A coding model may notice that no ArcGIS Pro version was specified. A research model may notice that “current” has no date. A document-extraction prompt may have no rule for illegible OCR. A Māori place-name task may have no distinction between official, historical and locally held names.
It is still slightly absurd. We have reached a point where a human can ask an AI to write the instructions for an AI, review those instructions, then give them to an AI and later ask another AI whether the first AI followed them. GIS practitioners who survived ModelBuilder should resist the urge to be smug.
The same discipline helps with writing. Telling AI to “make it engaging” often produces prose that sounds like every organisation has just embarked on an exciting new journey. For a Māori GIS audience, that problem can be worse if the model decorates generic corporate writing with a few Māori terms and mistakes that for cultural grounding. A paragraph can contain “rangatiratanga”, “kaitiakitanga” and “manaakitanga” and still recommend an architecture that removes practical control from the people whose data is involved.
Voice instructions need to be concrete. Tell the model the audience already understands GIS. Ask for plain New Zealand English, developed paragraphs, technical specificity and restrained use of te reo Māori where it is natural. Tell it not to manufacture whakataukī, cultural interpretations or personal experience. Tell it not to explain every Māori term as though the reader has arrived from another planet. Tell it not to turn every observation into a lesson about innovation. Then edit the result.
AI-generated prose has become recognisable less because of particular words than because of its behaviour. It introduces everything, arranges arguments into suspiciously neat shapes, restates the point after making it, and finishes sections by explaining that the preceding material was important. If asked for humour, it may also produce one polished joke per paragraph until the article reads as though the GIS team has hired a motivational speaker.
The cure is ordinary editing. Combine paragraphs that were split for effect. Delete transition sentences. Remove the second explanation. Replace vague claims with actual examples. Allow a paragraph to develop instead of forcing every idea onto its own line. If a statement is funny because it is true, do not add another sentence explaining the joke.
The answer still has to survive the map
Verification may be the most important part of this emerging skill because good AI is much more dangerous than obviously bad AI. An obviously poor model is irritating. A model that is correct most of the time earns trust, and its mistakes arrive surrounded by fluent, sensible material.
Code gives us a fairly direct test. Run it. Use a copy of the data. Check the installed library and GIS version. Confirm field names. Inspect counts before and after. Test nulls, empty input and odd values. Look for destructive writes. Check whether the coordinate reference system is preserved or deliberately transformed. If the AI recommends a package, confirm the package exists before redesigning the workflow around it.
Spatial facts need just as much discipline. Put the coordinate on a map. Check the scale. Confirm whether the source is current. For an ingoa wāhi, compare the name in the original source with the New Zealand Gazetteer where relevant, historical mapping, other records and locally held knowledge appropriate to the kaupapa. Keep official, historical and local names distinct instead of asking AI to select one “correct” version. The place-checking workflow exists for exactly this reason.
The distinction between finding and confirming is particularly important in Māori GIS. AI can identify a shape in LiDAR or imagery that resembles an archaeological form. That does not make the shape a pā. A model can find repeated references to a customary place in historical documents. That does not settle its exact location. It can compare a statutory boundary with several published rohe representations. It cannot decide which representation expresses local authority for a new kaupapa.
A useful GIS dataset can preserve those distinctions explicitly. model_detected, research_candidate, reviewed and confirmed are much safer concepts than one field called TRUE_FALSE. Ambiguity is information. Throwing it away because a map looks tidier is not quality assurance.
This is also why provenance matters more as AI becomes more capable. If an AI summarises a source, keep the source. If it extracts a candidate name, keep the page. If it proposes a match, keep the evidence used to review that match. If it generates code, keep the test result and version information. A fluent answer should not be allowed to sever the chain back to where the information came from.
The rule that AI output is a candidate finding and the source remains the evidence is deliberately conservative. It also scales well. The model may improve dramatically next month. The source will still be the source.
There is a final verification problem that is easy to overlook: maps produced as images by generative AI. They can look excellent while having no geographic integrity at all. Coastlines, labels, roads and boundaries can be plausible-looking decoration rather than geometry. For communication artwork that may be fine if nobody is claiming geographic accuracy. For analytical or decision-support mapping it is not GIS. Keep the real geometry, projection, source datasets and reproducible process.
Māori GIS is particularly poorly served by a beautiful fake map. A machine can invent an attractive approximation of Aotearoa, place a convincing cluster of Māori names around it and still be wrong in ways that matter. There is no prize for cartographic confidence.
The small syllabus that did not exist last year
The awkward part of learning all this is that some of the knowledge expires quickly. SQL remains SQL. NZTM2000 does not normally receive a surprise interface update. A QGIS expression written several years ago may need adjustment, but the concept has not been rebranded by three competing vendors in the same quarter.
AI product knowledge is different. Models appear, disappear and change position in the menu. Reasoning settings change. Context limits increase. Subscription allowances are revised. Coding assistants gain agent functions. Research tools acquire connectors. Employers approve one service and restrict another. A prompt technique that seemed essential six months ago may now be unnecessary because the model got better at following ordinary instructions.
That does not mean Māori GIS practitioners should become full-time followers of AI release notes. The industry can easily consume more attention than it returns. A short monthly review is enough for most people.
- Which model is now the sensible default for routine work, and which one is worth saving for difficult reasoning?
- Have usage limits, credits, reasoning settings or subscription rules changed?
- Has coding improved for the GIS tools you actually use, including Python, PyQGIS, ArcPy, SQL, Arcade or JavaScript?
- Have research, project, agent or file-handling modes changed enough to improve a real workflow?
- Has your organisation changed which AI services or connectors are approved?
- Are there new privacy, retention, data-residency or training-use conditions that matter to your work?
- Could a task involving Māori-held or sensitive information now be done locally, with synthetic data, or with less information leaving the current environment?
- Is a prompt ritual you learnt earlier still solving a real problem?
- What can safely be ignored for another month?
The last item deserves to stay. There is no professional requirement to learn every AI feature as it arrives. Knowing when not to adopt something is part of technical judgement. A new agent that can operate a browser autonomously is interesting. If your job this week is checking ten public Gazetteer records, it may not require a migration programme.
This is also where Māori data governance keeps the argument grounded. The arrival of a better model does not change who should have authority over Māori data. A longer context window does not create permission to upload more. A local model can reduce external data movement, but does not decide tikanga. An enterprise connector can make organisational files searchable, but access rights inside a system do not settle whether every permitted secondary use is appropriate. Te Kāhui Raraunga's Māori AI Governance Framework and the principles of Te Mana Raraunga are therefore relevant to everyday AI choices, not only to large national AI programmes.
The practical version is simple enough. Before using AI, understand the kaupapa and the information. Choose the smallest suitable boundary around the data. Give the system only the context it needs. Match the model and mode to the actual task. Ask for outputs that preserve uncertainty and provenance. Test code. Check places. Keep the authoritative source outside the conversation. Do not confuse fluency with authority.
That is quite a lot of judgement for a technology that still presents itself mainly as a text box.
It also fits Māori GIS rather well. Good Māori GIS has never been only about operating software. It asks where information came from, who has authority over it, what relationships are carried with it, what should be shared, what should remain close, and whether the map serves the kaupapa rather than simply proving that somebody knew how to make one. AI adds another technical system to those decisions, plus several new menus and a billing model nobody asked for.
The map may still be the easy part.
Related material
For practical local document search, see Local AI and Māori GIS. The wider sovereignty and governance treatment is in AI and Māori GIS and Māori Data Sovereignty for GIS. The earlier Ngā Poutama article Safe and sovereign AI for Māori GIS covers AI Haumaru and AI Motuhake in more detail.
Current product references
Product names, model availability and usage rules change quickly. These links were checked for this article on 30 August 2026.