Laying the foundation for generative AI and multi-agent systems in environmental assessment: building a curated dataset from the Danish EA Hub

The relevance and roles of artificial intelligence (AI) within impact assessment (IA) depends critically on the quality and relevance of the underlying data. This paper explores the development of a curated dataset of environmental assessment (EA) texts to support generative AI applications, including AI agents and modular multi-agent systems. Using the Danish EA Hub as a case study, we outline the key considerations involved in creating such a dataset, with particular attention to user needs, quality assurance, structuring, copyright and ownership, ethics and mechanisms for continuous updating. The curation process is analysed through a socio-technical lens, highlighting how data preparation is shaped by technical, legal, and institutional factors. The curated dataset ensures that AI systems are trained on context-specific, procedurally aligned, and legally compliant information – addressing the risk of relying on uncontrolled online sources. Finally, we outline the future potential of this dataset to support task-specific AI agents across various EA stages, from screening to compliance. The results highlight the foundational role of curated data in enabling responsible and effective AI integration in environmental governance.