This research paper examines how large language models (LLMs) can improve the classification and analysis of Canadian philanthropic grantmaking data. It focuses on the T3010 Registered Charity Information Return, which provides the most comprehensive administrative data on Canadian charities but has significant limitations, including reporting delays, inconsistent self-reported information, and categories that are either too broad or too detailed for meaningful analysis.
The paper proposes using LLMs to process charities’ textual descriptions of their activities and assign them to a more practical 13-category typology, covering areas such as education, health, social services, arts and culture, environment, and grantmaking. Rather than having an LLM directly classify the data, the approach uses LLMs to help develop replicable programs and language-processing methods for systematic classification.
Overall, the research explores the potential and methodological limitations of LLM-assisted analysis as a way to generate more detailed, reusable intelligence about philanthropic activity in Canada, with applications for researchers, foundations, policymakers, and sector organizations.
