Toward Calibrated Trust: What Working Scientists Need from an AI Co-Scientist

The Biology and Environmental Research (BER) Intelligence Layer (BERIL, https://beril-doe.org/) provides an AI agent interaction layer on top of the DOE Systems Biology Knowledgebase (KBase, kbase.us)’s new data lakehouse architecture. BERIL is an agentic co-scientist environment that scientists can use to discover data and carry out research projects. Success in this new mode of scientific work requires an understanding of how researchers will interact with agentic co-scientists. Here, we present findings from interviews with 11 BER scientists who used BERIL’s co-scientist to work with data in the KBase Lakehouse during a one-day workshop. Across the board, the agent's analytical capability was rarely the limiting factor; the harder challenge was for scientists to stay in command of the collaboration: maintaining a mental model of the co-scientist's process, judging what to trust, and deciding when to verify. BERIL's ability to access the KBase data lakehouse and integrate that data with other sources emerged as its key enabling strength. Because BER scientists already have established preferences for certain AI agents and skills, this capability is a key asset that distinguishes BERIL from other co-scientist systems and increases its effectiveness for researchers across the DOE landscape and beyond. Results also pointed to opportunities for supporting users’ mental models of BERIL’s actions and for helping them calibrate their trust in the co-scientist. Together, our findings indicate the importance of meeting users where they are, considering their needs and expectations as we introduce new tooling to accelerate their scientific workflows.