When and How AI Should Assist Brainstorming for AI Impact Assessment

TL;DR

This study explores how AI can effectively support team brainstorming in AI impact assessments, emphasizing appropriate timing, interaction style, and role to enhance collaboration and idea quality.

Contribution

The paper adapts and evaluates AI interventions in structured brainstorming methods, providing design guidance for AI support in collaborative impact assessment.

Findings

AI improved impact assessment quality for general-purpose AI but not for specialized applications.

AI should offer hints rather than solutions during early ideation phases.

Effective AI support involves facilitating idea structuring and leveraging team expertise.

Abstract

A key task in AI practice is to assess potential impacts to prevent harm. Current AI tools assisting AI impact assessment have not been designed or evaluated for collaborative team brainstorming, and they do not capture the range of views in diverse teams. We studied how AI can support team brainstorming during AI impact assessment and made three contributions. First, we adapted two structured methods from strategic foresight and co-designed AI interventions for them in five in-person workshops with 28 participants in total. Second, we evaluated the interventions in ten in-person workshops with 54 participants, finding that AI improved impact assessment quality and brainstorming perceptions for a general-purpose AI use (a chatbot companion) but not for a specialised one (a kidney allocation application). Third, our findings result in broader design guidance for AI assistance in brainstorming: AI should only offer hints and not solutions during early ideation, initiating interaction only when participants face fixation or saturation; it should facilitate structuring ideas during convergence; leverage expertise to refine ideas; and overall, it should serve more in support of tedious brainstorming process tasks, rather than ideation that teams value to do themselves.