GeoAI collapse? Ethical implications of synthetic geospatial data use

As synthetic data become increasingly embedded in GeoAI workflows, the long-term risks of recursive training on AI-generated inputs remain insufficiently understood. This study introduces and empirically examines the concept of GeoAI collapse–a degenerative process in which repeated reliance on synthetic geospatial data leads to progressive performance degradation. Focusing on semantic segmentation of street-level imagery, we simulated multigenerational training cycles using a conditional generative adversarial network pix2pix. Results demonstrated substantial declines in both visual fidelity and quantitative performance metrics, with rare place-based features exhibiting near-total collapse by later generations. These findings reveal structural vulnerabilities in GeoAI pipelines and highlight the ethical risks posed by unscreened, synthetic data entering public geospatial datasets. Beyond technical degradation, the study situates these risks within a broader phenomenon of digital placelessness–the erosion of geographic specificity and contextual meaning as AI-generated representations progressively abstract the lived realities of place. To address these challenges, we propose provenance-aware evaluation protocols that emphasize resilience, spatial fairness, and transparency. This work calls for a critical reframing of synthetic data practices in GeoAI to safeguard the integrity of geographic knowledge production.