Beyond the MAP: open audience questions

“It rained heavily in my neighbourhood last week. If it rains that heavily again next week, where could flooding occur?” This is the kind of question I want my mother-in-law to be able to ask a future geographic information system. It also captures much of what I wanted to bring to the Beyond the MAP panel in Sint-Niklaas on 10 September. She should be able to explain her problem without knowing which datasets, hydrological models or GIS operations are required. Yet answering it would demand considerable geographic knowledge: what counts as her neighbourhood, which rainfall event is being considered, how terrain and drainage affect the outcome, and how uncertain the prediction is. Making GIS easy to use does not make GIScience easy.

The appeal of natural-language GIS is that this complexity could become accessible through an ordinary conversation. The user describes the problem, and the system helps work out how to investigate it. The answer might still be a map: language is excellent for asking the question, while a map is often still excellent for understanding the answer. What changes is how we arrive there. Maps have evolved from paper to digital, interactive and dynamic representations, but mapping remains the process of selecting and transforming aspects of reality into something we can use. I see the next development as people and intelligent systems collaborating more closely in that process. Beyond the map does not mean beyond geography.

For that collaboration to work, geographic knowledge needs to shape the intelligence underneath. My background in moving objects and spatial reasoning, together with my experience of knowledge-based, data-driven and generative approaches, makes me particularly interested in how these can complement each other. Explicit knowledge can express constraints, learning methods can reveal patterns, and generative models can help interpret questions and connect tools. This is why I think in two directions: AI for Geo, but also Geo for AI. Geography can contribute to AI itself, because talking convincingly about the world does not automatically mean reasoning correctly about its spatial relationships. Research on world models explores how systems can learn from observations and interactions to predict how physical situations develop. This remains a research challenge, but sensing, topology, movement and change over time are problems our community has studied for decades.

That geographic competence becomes especially important as we move from generation to delegation. I sometimes write “aGENtic AI”, keeping GEN inside the word, because generating content remains useful while the scope of the system expands. With access to tools and procedures, agents can increasingly carry out work. In a network context, an agent could compare a field report with a digital twin, investigate a discrepancy and prepare an update for review. This is a direction to explore, with the degree of autonomy depending on what works reliably and what is at stake. Preparing an exploratory analysis and issuing an operational instruction carry different responsibilities. People consequently need to define goals, check evidence and decide when to intervene; domain expertise becomes central to supervising the work.

Delegation therefore asks us to think about the whole working environment. A GIS is not just software, and an AI system is not just a model. Both involve hardware, data, people and organisational arrangements that determine whether a capability becomes useful in practice. Reusable skills are interesting here because they can translate expertise into procedures: which sources to consult, which checks to perform and how to report a result. When experts identify shortcomings, their feedback can improve subsequent versions. Organisational learning can accumulate in a skill, provided that we deliberately maintain and validate it. This offers a way for knowledge developed through individual interactions to become available to colleagues and future workflows.

Those checks matter because fluent answers and successful tool execution do not by themselves establish that an analysis is valid. Garbage in, garbage out remains relevant, but good inputs alone are no guarantee either: a system can misunderstand the question or apply an unsuitable method to reliable data. Quality therefore needs to be judged in relation to the decision, using independent observations, established methods and explicit spatial constraints, with uncertainty kept visible. For critical infrastructure, I would put it simply: don’t let the model invent the network; let it interrogate the trusted network. Delegating work should make responsibility clearer, including who decides whether the result is suitable for use.

Once we consider the whole system, strategic questions follow. Who controls the interface through which people ask questions, who provides trusted data, and who coordinates the workflow? These roles may be distributed across established GIS providers, specialist GeoAI organisations and general AI platforms. I would not predict a winner, but these dependencies deserve attention when deciding what expertise and infrastructure an organisation should develop. They also make me question the word “transition”. A transition to what? If continuous change becomes the normal condition, organisations need a lasting capacity to adapt. That means broad AI literacy alongside people who actively follow developments, test their relevance and translate them into practice.

For me, this brings the discussion back to the question I would like my mother-in-law to ask. Productive collaboration depends on combining human purposes, experience and judgement with systems that can help find information, explore alternatives and carry out work. My advice is to become very good at your domain, and very good at using AI within that domain. Learning how to bring those competences together also motivates my involvement as one of the senior experts helping build AI Community@UGent. We share developments in GenAI techniques and almost daily practical examples and best practices. People within and outside UGent are welcome to follow or contribute. If you are interested, send me an email and I can give you access.