Program on the Environment

NOAA Fisheries is developing a public e-library on salmon-stressor response functions (SRFs) to support conservation and management of salmon populations facing increasing environmental stressors such as climate change and habitat degradation. However, budget cuts and loss of personnel have constrained the speed and scale of data extraction from scientific literature. The purpose of this study was to evaluate how well Google Gemini can extract, summarize, and interpret salmon SRF data from published research papers. During my internship with NOAA Fisheries, I assessed Gemini’s performance using a test set of 12 pre-existing SRF entries with corresponding research papers and manually extracted metadata. I compared outputs generated by Gemini (using both an unguided and a guided prompt) to human extracted metadata, and evaluated extraction accuracy while also measuring consistency of results across different Gemini Pro accounts. Results indicate that the guided prompt substantially improved the correctness of extracted metadata compared to the unguided prompt, which suggests that prompt design plays a part in Gemini’s ability to extract SRF metadata from scientific literature. Additionally, the replicability test showed that the guided prompt produced reasonably consistent results, but variability remained. These findings suggest that Gemini showed promise as a workflow support tool when paired with carefully designed prompts. The most effective application is AI-assisted extraction, in which Gemini helps identify likely answers and reduces the time required for human reviewers to verify and complete SRF entries. This approach could improve efficiency and scalability in building the salmon SRF e-library despite NOAA’s ongoing resource constraints.