AIRUS: a simple workflow for AI-assisted exploration of scientific data

The development "reasoning" large language models (LLMs) seems poised to transform data analysis in all fields of science. This note describes a simple, iterative workflow termed AIRUS (AI Research Under Supervision), that allows working scientists, including those who lack extensive AI or programming expertise, to immediately start taking advantage of these capabilities. In the proposed workflow, a researcher uses an LLM to generate hypotheses, produce code, interpret results, and refine its approach in repeated cycles, only intervening when necessary. This can be run by simple cut-and-paste of code and results between web-based LLMs and an online Jupyter notebook. We demonstrate the workflow with two examples: one analysis of synthetic data, and one analysis of data from the International Brain Lab.