A shorter path to an experiment
Research involves more than fitting a model. It includes reading documentation, comparing definitions, preparing data and translating a question into code. AI can assist with these tasks, giving a researcher more ways to explore a problem.
The resulting code or explanation is a starting point. A persuasive answer does not establish that a dataset was available at the time of a decision, that an implementation is correct, or that a relationship will persist.
Make the claim testable
NIST’s generative AI profile identifies confabulation as a risk: a system can produce content that is wrong while presenting it confidently. This matters when an assistant proposes an explanation or summarizes a source. [1]
Keep the evidence visible
Our preferred workflow is to write the question first, identify the evidence needed to answer it, and make each computational step inspectable. For example, an AI-generated data transformation should be checked against a small case with a known answer before it is used in a larger experiment.
The value of AI lies in expanding what can be investigated. The research conclusion still belongs to the experiment: its data, assumptions, results and limits.