The Situate AI Guidebook: Co-Designing a Toolkit to Support Multi-Stakeholder Early-stage Deliberations Around Public Sector AI Proposals

Early-stage decision-making is critical: The paper finds that AI failures in public sector applications often originate from decisions made during initial ideation and design phases, particularly in problem formulation, suggesting that intervention at these early stages could prevent downstream harms in consequential domains like child welfare and criminal justice.

Multi-stakeholder input is essential: The Situate AI Guidebook was developed through collaboration with agency leaders, AI developers, frontline workers, and community advocates, reflecting the finding that responsible AI deployment requires diverse perspectives beyond technical teams alone.

Structured deliberation processes are needed: The guidebook provides a systematic framework organized around four key areas—goals and intended use, societal and legal considerations, data and modeling constraints, and organizational governance—addressing a gap where public sector agencies currently lack standardized processes for evaluating whether to proceed with AI projects.

Current AI deployment practices are insufficient: The paper documents that AI systems are "rapidly deploying" in high-stakes public sector decisions despite evidence that they "often fail to improve services in practice," indicating that existing oversight and decision-making processes are inadequate for protecting public welfare.

Practical toolkit design matters for adoption: By centering the guidebook's design on the actual challenges, needs, and desires expressed by public sector stakeholders themselves, the research suggests that responsible AI tools are more likely to be adopted and effective when co-designed with the communities that will use them.