AI Design is a semi-automated framework that guides stakeholders in populating impact assessment reports. It consists of two components. The first component, StakeLinker (A), gathers input from various stakeholders regarding different aspects of the AI system. This is achieved through 32 statements that are based on existing literature. The second component, an LLM-powered tool, called FillGen (B), processes stakeholders' input. It summarizes the input and generates new information about the system, covering system’s usse (SysInfoGen), risks (RiskGen), benefits (BenefitGen), and mitigations (MitigationGen).

The content from FillGen is subsequently used to pre-populate an impact assessment report (C). The stakeholders review the pre-populated report, and present it to regulatory experts for final approval.

For generalizability purposes, we provide FillGen's code and pre-populated reports for two representative AI systems (i.e., a chatbot assessing responses to vocational training tasks, and a system assessing damage after natural disasters)

.