A Framework for Reproducible AI-Assisted Research in the Social and Behavioral Sciences

Artificial Intelligence (AI) is increasingly used in social and behavioral research for tasks including literature search, stimulus generation, response annotation, code generation, and writing. These uses challenge conventional understandings of reproducibility because AI-assisted procedures may shape not only the final analysis but also the empirical materials, data-processing steps, and analytical inputs on which reported findings depend. We propose a conceptual framework for deciding when, what, and how to document AI-assisted research to support reproducibility. It distinguishes between AI as a research partner, where transparent disclosure may often be sufficient, and AI as a methodological instrument, where AI outputs enter the evidential chain and therefore require more detailed documentation. We structure seven guiding questions across three stages: (a) clarifying the level of reproducibility required for AI use, (b) determining what needs to be reproducible, and (c) deciding how AI assistance should be documented in manuscripts and supplementary materials. Examples from AI-assisted research demonstrate how these questions guide documentation decisions across different reproducibility requirements. Overall, the framework shifts the focus from generic AI disclosure to role-based documentation that makes AI-assisted workflows understandable, auditable, and, where necessary, reproducible. It also underscores that reproducibility should not be reduced to reporting prompts and AI model settings alone, but requires attention to output variability, changes introduced through human oversight, and the infrastructures through which AI systems are accessed, versioned, and preserved.