Human-Computer Interaction Ph.D. Thesis Defense - Frederic Gmeiner

 

Despite Generative AI's potential to support complex, open-ended creative problem-solving and analytical tasks, such as designing multi-objective mechanical parts or analyzing data, professionals often struggle to integrate these systems into their workflows effectively. Key challenges include misalignment between outputs and intentions, uncertainty about how to guide the system, and reduced cognitive engagement when tasks are overly delegated to automation—often resulting in insufficient problem exploration and limited ability to evaluate generated outcomes. I argue that to better utilize GenAI's capabilities for supporting complex open-ended creative problem-solving tasks, we need not only to improve AI models, but also to design new interactions and systems that better align with and enhance human thought processes.

My research addresses this challenge by rethinking human-AI interaction for cognitively demanding professional workflows. I study how people use GenAI in AI-assisted domains such as mechanical design, slide presentation authoring, and data analysis, and develop interaction techniques that scaffold reflection, sharpen problem formulation, and support deliberate cognitive engagement where human judgment is critical. Building on cognitive and learning science principles, such as metacognitive thinking and Socratic dialogue, and on design theory concepts, such as reflection-in-action, this thesis explores how GenAI systems can better align with and enhance human thought processes.

In the first research thread, I examine how mechanical and architectural designers engage with novel AI-based design tools in multi-objective manufacturing tasks. From these studies, I identify support strategies—such as reflective questioning and thought externalization—that improve designers’ cognitive engagement and outcomes. I further demonstrate how personal voice-based agents can facilitate cognitive support in GenAI workflows, leading to higher-quality outcomes. In a second thread, I develop and evaluate Intent Tagging, an input technique that supports granular and reflective control of Large Language Models during slide presentation authoring through lightweight, structured, and adaptive graphical interface elements. Finally, through MindKicks, I introduce a modular plug-in architecture for augmenting existing AI-assisted workflows with proactive cognitive “sidekicks.” Through a study with data analysts, I show how a MindKicks-based support system can complement chat-based coding agents, increase analytical AI-engagement, and help analysts discover more data issues. In sum, this thesis contributes interaction techniques, systems, prototyping methods, and empirical insights to advance human-AI interaction toward enabling people to think more deeply, create more thoughtfully, and solve more complex problems with the help of AI.