From Geospatial Data Gaps to AI Representativeness Bias: A Framework for Low-Data Economies
13/11/2026 –, Sala de Video

Across much of Sub-Saharan Africa and South Asia, formal land registration covers only a small fraction of land parcels in rural and informally held areas. Nigeria's own government places its national formal titling rate below 5 per cent (Federal Ministry of Housing and Urban Development, Nigeria, 2025). The wealth is real. The records are not. This is the condition under which AI systems are increasingly trained and deployed, and it has received far less attention than the kind of bias the AI fairness literature usually worries about.

That literature is good at catching unequal treatment. Demographic parity, equalised odds, calibration: these tools ask whether identifiable groups are treated differently by an algorithm. But a fairness audit built on these metrics would not detect the pattern this paper is concerned with, because the affected populations are absent from the training records altogether, not merely treated differently within them. No model output exists to audit, because no input data exists to audit it with.

This paper argues that geospatial data absence, rather than discriminatory model design, is a primary driver of AI representativeness bias in low-data economies. To make this argument tractable, the paper develops the Geospatial-AI Alignment (GAIA-M) framework and operationalises it through the GAIA-M Data Infrastructure Quality Index (GAIA-M DIQI), a composite score across five dimensions: cadastral coverage, data accessibility, temporal stability, GIS layer availability, and informal tenure recognition.

GAIA-M DIQI = (0.30 × CC) + (0.20 × DA) + (0.20 × TS) + (0.15 × GIS) + (0.15 × ITR)

Each dimension is scored from 1 to 10 against a rubric grounded in ISO 19157 (ISO, 2013) and the UN-Habitat Fit-for-Purpose Land Administration framework (Enemark, 2009), sorting territories into four risk bands from Very High (0–39) to Low (80–100). Most geospatial data quality work asks whether a map is accurate, complete, and consistent. GAIA-M asks a different question on top of that: whether the data is good enough to make an AI system trained on it fair. The weighting structure reflects the centrality of cadastral visibility to AI training data, with cadastral coverage assigned the highest weight; sensitivity testing using equal weighting and three alternative weighting schemes produced the same risk classification for every country considered.

Nigeria, Bangladesh, and Indonesia were chosen deliberately for how different their cadastral systems are from one another, following a most-different systems design (George and Bennett, 2005). All three share large Muslim populations and well-documented land administration problems, which is what makes the contrast between them meaningful. Rwanda is not a Muslim-majority country, and was not chosen by the same logic; it serves a different purpose, set out below.

Each number used in this comparison is drawn from a named, publicly available source, not estimated or constructed by the authors. Nigeria sits below 5 per cent formal registration, a case of what the framework calls static absence, records that were largely never created, rather than records that have since gone bad. Bangladesh sits at roughly 50 per cent, with about half of all land parcels held through unregistered documents according to a 2020 household survey (Rahman and Hossain, 2020). But that figure understates the deeper problem, which is dynamic rather than static: land along the Brahmaputra, Ganges, and Meghna shifts fast enough that a registered boundary can be substantially wrong within months. Indonesia has reached approximately 71.5 per cent of total land area through more than a decade of systematic registration under its PTSL programme (Journal of Property Research, 2025), the strongest of the three Muslim-majority cases. And then there is Rwanda, which has achieved near-universal registration, more than 10 million parcels titled nationwide between 2009 and 2013 (Gillingham and Buckle, 2014), despite having the lowest GDP per capita of the four countries, roughly 1,070 US dollars (World Bank, 2024).

That last fact is what makes Rwanda worth including. It breaks an obvious alternative explanation: that the whole pattern is really just about national income, not land administration at all. If income were driving the result, Rwanda should be the weakest case here. It is the strongest. One country is not proof of anything, but it is real, sourced evidence against the simpler explanation.

Nigeria scores 28, Bangladesh 46, Indonesia 68, and Rwanda 72. The full dimension-level breakdown behind each of those numbers is set out in the accompanying description.

Rwanda's composite score is more cautious than its registration figure alone would suggest. Cadastral coverage is well evidenced; accessibility, temporal stability, and GIS infrastructure maturity are not, so the score holds back where the evidence does. No correlation coefficient, significance test, or spatial autocorrelation statistic appears anywhere in this paper. Four countries is not a sample large enough to carry that kind of weight, and the comparison is offered as illustrative throughout, not as something statistically tested.

For GIScience, the contribution is this: spatial data quality assessment becomes a direct input into algorithmic governance, through a scoring instrument that is transparent and recomputable rather than an opaque statistical model. For AI governance, the contribution is a pre-deployment index that names the type of representativeness risk, absence, decay, or concentration, before a model is trained rather than after it has already failed. The next step is to test this properly: Pearson correlation and spatial autocorrelation analysis would be the right tools, once sub-national data exists to run them on. Indonesia is the strongest candidate for that work, since province-level religion data is already published by BPS-Statistics Indonesia, and province-level land registration progress is already tracked by the national land agency, ATR/BPN. The other three countries in this paper do not yet have both halves of that data publicly available; Indonesia does.

Ayobami Oluyemi Ayoola is currently completing an MSc in Islamic Finance and AI for Business at the University of Dundee, with a background in Surveying and Geoinformatics that grounds this work. His work sits at the intersection of geospatial data infrastructure and artificial intelligence, looking at how gaps in cadastral and land administration systems shape AI governance and financial inclusion in low-data economies.

He developed the GAIA-M framework to identify how deficiencies in geospatial data quietly turn into representativeness bias in the AI systems used for credit, land, and financial decisions. Because he has worked directly with cadastral systems, the questions he asks about AI fairness stay tied to how land data actually gets produced and where institutional constraints get in the way, rather than staying purely theoretical.

His current work compares Nigeria, Bangladesh, Indonesia, and Rwanda, asking whether geospatial data quality assessment can flag risk before an AI system is deployed, rather than after exclusionary outcomes have already happened.