Augmenting Expert Cognition in the Age of Generative AI: Insights from Document-Centric Knowledge Work

TL;DR

The paper addresses how Generative AI influences expert cognition in document-centric knowledge work and seeks to preserve expertise while leveraging automation. It employs two empirical studies—business sensemaking with an AI-assisted system and scholarly survey authoring—to observe patterns of AI delegation and information verification by domain experts. Key findings show that experts welcome AI for routine information foraging but insist on retaining control over synthesis and interpretation, and that verification and metacognitive supports are essential. The work provides design implications for GenAI systems to selectively delegate, verify outputs, and support expertise development, aiming to augment rather than diminish professional judgment in real-world workflows.

Abstract

As Generative AI (GenAI) capabilities expand, understanding how to preserve and develop human expertise while leveraging AI's benefits becomes increasingly critical. Through empirical studies in two contexts -- survey article authoring in scholarly research and business document sensemaking -- we examine how domain expertise shapes patterns of AI delegation and information processing among knowledge workers. Our findings reveal that while experts welcome AI assistance with repetitive information foraging tasks, they prefer to retain control over complex synthesis and interpretation activities that require nuanced domain understanding. We identify implications for designing GenAI systems that support expert cognition. These include enabling selective delegation aligned with expertise levels, preserving expert agency over critical analytical tasks, considering varying levels of domain expertise in system design, and supporting verification mechanisms that help users calibrate their reliance while deepening expertise. We discuss the inherent tension between reducing cognitive load through automation and maintaining the deliberate practice necessary for expertise development. Lastly, we suggest approaches for designing systems that provide metacognitive support, moving beyond simple task automation toward actively supporting expertise development. This work contributes to our understanding of how to design AI systems that augment rather than diminish human expertise in document-centric workflows.