Escaping the Delphic Trap: Providing Variation Affordances to Foster Agency and Resilience in AI-Mediated Sensemaking

Abstract

With emergent capabilities of generative AI, many systems have been eager to adopt features providing \textit{synthesis affordances}: properties that enable and entice users in engaging with AI-synthesized information. There are contexts where offering synthesis affordances may be appropriate---perhaps, even useful. However, approaches taken by popular design patterns in providing them are often not \textit{resilient}. They fail upon breakdowns stemming from AI limitations, human factors, or information quality issues. More concerningly, these design patterns distribute agency inappropriately between humans and AI by granting excessive agency to AI systems. Such configurations lure users into what we term the ``Delphic Trap,'' where users are enticed to satisfice to suboptimal information practice. Drawing on theories from cognitive and ecological psychology and work in Human-Computer Interaction, we conceptualize \textit{variation affordances}, defining them as system properties that invite users to engage with the inherent variation within information collections. Systems can offer these affordances through steerable controls that support and invite users in engaging in productive friction during information seeking and sensemaking. We argue that a way to instantiate a more appropriate delegation of agency between humans and AI systems, in this context, is through providing variation affordances; doing so may help users engage in more intentional information actions. We design and evaluate several ways to provide these affordances. Chapter 1 establishes the motivating background for this work. Chapter 2 reviews structure-mapping theory literature to inform approaches for helping users utilize variation. Chapter 3 presents an eye-tracking ablation study (n=24) examining how participants interact with different feature sets providing various variation affordances, discussing functional aspects to consider when implementing these affordances. Chapter 4 examines how common design patterns offering synthesis affordances lack resilience and may cause users to fall into the Delphic Trap. Addressing these risks, we propose a design intervention providing variation affordances through ``AI Highlighters''. Chapter 5 presents a formative study (n=24) assessing user interactions with our design intervention and its variation affordances.