Exploring the Potential of Metacognitive Support Agents for Human-AI Co-Creation Conference: Designing Interactive Systems Conference; July 5–9, 2025; Funchal, Portugal Designing Interactive Systems Conference (DIS ’25), July 5–9, 2025, Funchal, Portugal DOI: 10.1145/3715336.3735785 ISBN: 979-8-4007
Frederic Gmeiner, Kaitao Luo, Ye Wang, Kenneth Holstein, Nikolas Martelaro · 2025
Read Exploring the Potential of Metacognitive Support Agents for Human-AI Co-Creation Conference: Designing Interactive Systems Conference; July 5–9, 2025; Funchal, Portugal Designing Interactive Systems Conference (DIS ’25), July 5–9, 2025, Funchal, Portugal DOI: 10.1145/3715336.3735785 ISBN: 979-8-4007 by Frederic Gmeiner, Kaitao Luo, Ye Wang, Kenneth Holstein, Nikolas Martelaro (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-12879.txt.
Introduction
Frederic Gmeiner, Kaitao Luo, Ye Wang, Kenneth Holstein, and Nikolas Martelaro ask how AI agents could help designers think more reflectively, in a two thousand twenty-five conference paper. Designers using generative AI tools often struggle to state their criteria clearly and stay mentally engaged.
The team proposes metacognitive support agents that counter intent-formulation problems and cognitive offloading. They tested the idea through a Wizard of Oz study with twenty mechanical designers trying several support plans. Backed users made more feasible designs than unsupported users, though plans differed in effect. The paper weighs opportunities and trade-offs for future AI design tools.
Catalog record
| CREATOR | Frederic Gmeiner, Kaitao Luo, Ye Wang, Kenneth Holstein, Nikolas Martelaro |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 115 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-12879
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Exploring the Potential of Metacognitive Support Agents for Human-AI Co-Creation Conference: Designing Interactive Systems Conference; July 5–9, 2025; Funchal, Portugal Designing Interactive Systems Conference (DIS ’25), July 5–9, 2025, Funchal, Portugal DOI: 10.1145/3715336.3735785 ISBN: 979-8-4007-1485-6/2025/07 CCS: Human-centered computing Empirical studies in HCI CCS: Computing methodologies Artificial intelligence Frederic Gmeiner, Kaitao Luo, Ye Wang, Kenneth Holstein, Nikolas Martelaro Abstract . Despite the potential of generative AI (GenAI) design tools to enhance design processes, professionals often struggle to integrate AI into their workflows. Fundamental cognitive challenges include the need to specify all design criteria as distinct parameters upfront (intent formulation) and designers’ reduced cognitive involvement in the design process due to cognitive offloading, which can lead to insufficient problem exploration, underspecification, and limited ability to evaluate outcomes. Motivated by these challenges, we envision novel metacognitive support agents that assist designers in working more reflectively with GenAI. To explore this vision, we conducted exploratory prototyping through a Wizard of Oz elicitation study with 20 mechanical designers probing multiple metacognitive support strategies. We found that agent-supported users created more feasible designs than non-supported users, with differing impacts between support strategies. Based on these findings, we discuss opportunities and tradeoffs of metacognitive support agents and considerations for future AI-based design tools. 1. Introduction Generative AI (GenAI) models offer increasing capabilities in supporting design workflows by generating images ( Saharia et al. 2022 ) , videos ( Ho et al. 2022 ) , or complex mechanical parts ( formlabs 2020 ; Yang et al. 2020 ) . In mechanical design, working with AI allows designers to co-create designs that would be extremely tedious or even infeasible without AI support, such as reducing the weight of an electric wheelchair component ( formlabs 2020 ) or generating parts using emerging manufacturing processes ( Yang et al. 2020 ) with 3D geometric GenAI solvers. However, despite the growing promise of AI design tools to augment design processes, professionals often struggle to effectively integrate AI into their workflows ( Gmeiner et al. 2023 ; Zhang et al. 2021 ) . GenAI demands new (computational) workflows that require designers to work differently than they are used to or trained in ( Woodbury 2010 ; Menges and Ahlquist 2011 ; Gmeiner et al. 2023 ; Tankelevitch et al. 2023 ) . Current GenAI-supported workflows pose a set of unique cognitive challenges, including: (1) Intent Formulation : Designers have to specify all design criteria necessary for generating feasible parts as distinct parameters upfront instead of iteratively modeling, testing, and visualizing a part’s 3D geometry ( Gmeiner et al. 2023 ; Subramonyam et al. 2024 ) . This is a particular challenge for GenAI systems with lengthy and expensive inference times, such as complex 3D geometric solvers (e.g., ( Autodesk 2020 ) ). (2) Problem Exploration : To tackle design problems sufficiently, designers must thoroughly think through and consider many aspects, but GenAI workflow automation can reduce cognitive engagement and foster overreliance due to "cognitive offloading," making problem exploration more challenging ( Zamfirescu-Pereira et al. 2023 ; Zhang et al. 2021 ; Lee and See 2004 ) . (3) Outcome Evaluation : Designers are also required to evaluate generated designs according to the problem, but when their problem understanding is limited due to cognitive offloading, they won’t be able to effectively evaluate and refine generated designs ( Zhang et al. 2021 ; Tankelevitch et al. 2023 ) . Motivated by such cognitive challenges, we explore interaction patterns and interfaces to support professionals in more effectively working with AI-driven design tools. In this work, we follow an exploratory prototyping approach ( Zamfirescu-Pereira et al. 2021 ) to explore the potential of voice-based agents that support designers’ metacognition ( Flavell 1979 ) 1 1 1 Metacognition refers to mental processes of thinking about one’s own thinking, enabling individuals to regulate and improve their cognitive strategies by reflecting on their decisions and problem-solving approaches. while working on a manufacturing design task in a 3D GenAI CAD tool, where the designer specifies their goals as parameters and geometry within a graphical CAD interface. Broadly, we ask: What interfaces and interaction patterns can support designers in better thinking through and formulating design problems, and evaluating generated outcomes, when working with GenAI-based design tools? Informed by theories and findings from human-AI interaction, learning sciences, and the study of design processes, we engaged in exploratory prototyping ( Zamfirescu-Pereira et al. 2021 ) to explore a design space for metacognitive support agents through three different design probes ( Boehner et al. 2007 ) and observe how each influences designers’ processes and outcomes in a GenAI-based design task. In this prototyping process, we used the “Wizard of Oz” (WoZ) technique ( Dahlbäck et al. 1993 ; Zamfirescu-Pereira et al. 2021 ) , in which a human operator controlled the agent probes in a flexible manner but within certain probe-dependent constraints. Each probe followed a different support approach: (1) SocratAIs asks reflective questions to prompt deeper reflection-in-action and (2) HephAIstus prompts task planning and diagramming supported by suggestions for design strategies and software operation. While the first author enacted these two agents, we also included (3) external experts in mechanical and generative design from Autodesk to act as wizards in some sessions, who we invited to provide their own interpretations of metacognitive support strategies in a freeform manner. Since CAD-based work is highly visual-spatial, we explore all support agents through voice modality to reduce cognitive load. Inspired by the concept of "think-aloud computing" ( Krosnick et al. 2021 ) , we prompt designers to verbalize their thoughts while working on the design task to foster deeper reflection-in-action and to elicit their knowledge and situational intentions for the support agents. We conducted a formative study with 20 trained mechanical engineers new to working with generative AI systems. The designers were supported by one agent probe (or received no support in a control condition) while working on a realistic mechanical design task in the "Generative Design" extension of the commercial CAD software Autodesk Fusion 360 ( Autodesk 2020 ) . By comparing the design processes, outcome quality, and participant post-task interviews through video interaction and thematic analysis, we investigate the following research questions: RQ1 How do different agent support strategies impact the design process? RQ2 What are the perceived benefits and challenges of metacognitive support agents? Overall, we found that agent-supported users created more feasible designs than unsupported users. Most users actively engaged with and appreciated the agent’s support in helping them think through the design task and operate the software. We also identified that different agent strategies had different impacts. For example, question-asking strategies that prompted mental simulations or visualizations through sketching helped designers with intent formulation and problem exploration regarding the part’s mechanical loads. However, we also observed that asking questions alone was less impactful when users had solidified incorrect assumptions, and that sometimes agent support could lead to additional overreliance. Finally, our findings provide insight into design trade-offs and differences in user preferences for metacognitive support agents. We conclude by discussing design implications for future metacognitive support agents for GenAI-based design tools. While our paper explores support for mechanical design tasks, we discuss how our findings may generalize to other GenAI design activities.