AI as Expert-Assistive Tools

Abstract

While AI capabilities continue to grow, one of the chief questions that should be discussed is how to apply this increasingly powerful tool to assist human users. Particularly, we would like to answer the following research question: \textbf{How can we utilize and understand AI to create expert-assistive tools?} We consider the following two categories: creation/application of tools (utilization) and evaluation of tools (understanding). In the domain utilization we present two chapters: the application of deep learning based action recognition models to provide enhanced data categorization and analysis for basketball shot types, and string grammars and automated space exploration and design assembly to provide enhanced design speed for the construction of unmanned aerial vehicles (UAVs). In the domain of understanding we present three chapters: we evaluate the performance of GPT-3.5 and GPT-4 models on varied question answering tasks utilizing varied persona-based prompting methods, we assess the efficacy of LLM-based emails, both malign and benign, to be considered harmless by undergraduate users, and analyze the reasoning capabilities of flagship LLMs in the context of chess, deriving agentic design principles from our findings.