Inspectable AI for science: Governing generative AI with the Research Object approach

The authors propose the AI-RO framework, which governs interactions with generative AI in scientific research as inspectable structured components. The framework draws on Research Object and the FAIR principles, using interaction logs and metadata to record model configurations, prompts, and outputs, which is particularly critical for security and privacy (S&P) research. Its lightweight writing pipeline demonstrates how to have a language model synthesize manually written literature review notes and generate verifiable provenance records.