Geography according to ChatGPT

We invite you to an upcoming (online) kick-off session exploring how foundation models such as the large language model GPT or the text-to-image model Stable Diffusion describe and depict geographic space, categorize geographic features, perform in spatial reasoning, and implicitly apply principles of (geo)spatial data analysis. In 2019, we asked provocatively whether our community will be able to develop a chatbot-style GeoAI system that can replace a junior GIS Analyst by 2030; what are the social and technical implications if we get there by 2025?

This meeting will be the first in a series of activities in the tradition of Geographic Information Observatories (GIO) to develop a better understanding of the current abilities and limitations of these models and also serve as a starting point for discussing the design of the geo-foundation models of the near future. Successfully developing such models will require understanding why existing models struggle with certain types of inferences, e.g., topological reasoning, while returning strong results in other areas. 

The session(s) will combine technical discussions with those related to foundational principles of GeoAI, including themes such as neutrality and bias and how they already affect everyday interaction with systems such as ChatGPT and Stable Diffusion. We will also explore how to incorporate integrated data sources, e.g., knowledge graphs, to infuse facts into these models, reducing their tendency to hallucinate. Discussions will include lessons learned and how to apply them to fields such as Spatial Data Science, GeoAI, GIScience, and Geography, but also how our contributions inform the broader AI community, e.g., in terms of measuring geo-diversity in training data.