A workflow for human-centered machine-assisted hypothesis generation: comment on Banker et al. (2023)

Abstract

Large language models (LLMs) have the potential to revolutionize a key aspect of the scientific process—hypothesis generation. Banker et al. (2024) investigate how GPT-3 and GPT-4 can be used to generate novel hypotheses useful for social psychologists. Although timely, we argue that their approach overlooks the limitations of both humans and LLMs and does not incorporate crucial information on the inquiring researcher’s inner world (e.g., values, goals) and outer world (e.g., existing literature) into the hypothesis generation process. Instead, we propose a human-centered workflow (Hope et al., 2023) that recognizes the limitations and capabilities of both the researchers and LLMs. Our workflow features a process of iterative engagement between researchers and GPT-4 that augments—rather than displaces—each researcher’s unique role in the hypothesis generation process. (PsycInfo Database Record (c) 2024 APA, all rights reserved)