I’m thrilled to announce that our paper on AI Safety for Geospatial AI has been accepted to ACL Findings! 🎉
Yesterday, I had the privilege of presenting our work virtually at the AACL-IJCNLP 2025 conference in Mumbai. As LLMs become increasingly integrated into critical applications, AI safety has never been more important — especially when handling sensitive geospatial data and mitigating the associated risks.
What we built
Our contribution introduces three artifacts:
- GeoSAFE Taxonomy — a comprehensive framework for categorizing geospatial AI safety concerns, organized into six critical hazard categories (Environmental Hazards, Urban Safety, Public Health & Well-Being, Infrastructure Risk, Agriculture & Land Use, and Privacy & Surveillance).
- GeoSAFE Dataset — a curated benchmark of 12,078 LLM-human interaction prompts and responses, annotated for safety evaluation, built with data-centric AI principles to remove bias and improve fairness.
- GeoSAFEGuard — a fine-tuned moderation model (built on LlamaGuard and NeMo Guardrails) that achieves a 97% F1 score and significantly reduces false-safe rates in geospatial AI prompts.
We evaluated our model against multiple general-purpose models and demonstrated significant improvements in handling geospatial data risks.
Collaboration
This research was conducted in collaboration with the NASA-IMPACT team. I’m deeply grateful to all my co-authors — Manil Maskey, Shaona Ghosh, Rajat Shinde, Sumit Nawathe, and William Seawright — for their invaluable support, guidance, and collaborative spirit throughout this journey.
Looking forward
Excited to see how this work contributes to building more responsible and secure geospatial AI systems. 🌍🛡️