Ceci N'est Pas un Drone: Investigating the Impact of Design Representation on Design Decision Making When Using GenAI

TL;DR

This paper empirically investigates how design representation modalities—visual renderings, numerical performance data, or both—affect engineers' and designers' decisions when selecting from AI-generated UAV designs. Using two within-subject studies with drone hobbyists and STEM students, it shows that presenting only numerical performance data yields the highest accuracy in identifying Pareto-TOPSIS-optimal designs, while visual renderings can bias toward conventional, axisymmetric designs and reduce optimality detection. The work illuminates how cognitive load from many options and modality-driven biases shape human-AI collaboration in design exploration, offering concrete recommendations for tool design (e.g., option limits, modality toggles, and bias-mitigation workflows). Overall, the findings advance understanding of how design representations interact with GenAI-generated design spaces and have practical implications for deploying AI-assisted engineering design tools.

Abstract

With generative AI-powered design tools, designers and engineers can efficiently generate large numbers of design ideas. However, efficient exploration of these ideas requires designers to select a smaller group of potential solutions for further development. Therefore, the ability to judge and evaluate designs is critical for the successful use of generative design tools. Different design representation modalities can potentially affect designers' judgments. This work investigates how different design modalities, including visual rendering, numerical performance data, and a combination of both, affect designers' design selections from AI-generated design concepts for Uncrewed Aerial Vehicles. We found that different design modalities do affect designers' choices. Unexpectedly, we found that providing only numerical design performance data can lead to the best ability to select optimal designs. We also found that participants prefer visually conventional designs with axis-symmetry. The findings of this work provide insights into the interaction between human users and generative design systems.

Figures (15)

• Figure 1: Examples of design modalities used in the studies.
• Figure 2: Study 1: Participants' choice of design for consecutive questions in three design problems. V is with visual rendering, D is with numerical performance data, and M is with both visual rendering and numerical performance data.
• Figure 3: Study 1: Participants' choice for conventional design vs unusual design in percentage.
• Figure 4: Study 1: Participants' accuracy on selecting the optimal designs. Error bars represent a 95% confidence interval. p-value annotation legend: ns: 1.70e-02 < p <= 1.00e+00; *: 1.00e-02 < p <= 1.70e-02; **: 1.00e-03 < p <= 1.00e-02; ***: 1.00e-04 < p <= 1.00e-03; ****: p <= 1.00e-04. Participants are drone hobbyists.
• Figure 5: Study 2: Participants' choice of design for consecutive questions in three design problems. V is with visual rendering, D is with numerical performance data, and M is with both visual rendering and numerical performance data.