Dark Patterns as Quality Risks in GPT-Generated Geospatial Interface Prototypes

Abstract

Abstract. Generative AI tools are changing how interface ideas are turned into executable prototypes. While this can accelerate early design exploration, it also raises a quality concern: design goals expressed in prompts may be carried into the resulting interface in ways that affect user autonomy. This paper examines this issue in map based user interfaces. We generated a dataset of 90 front end prototypes representing a map based local search and navigation application, evaluating three of its core features: restaurant search, hotel choice, and route selection. For each feature, the prompts varied the intended design orientation, ranging from neutral decision support to commercial prioritization and transparency oriented assistance. Two reviewers inspected the generated prototypes for dark pattern mechanisms using categories from prior literature. A notable pattern emerged in the commercial priority condition, where both reviewers flagged all 30 prototypes as presenting potential dark pattern concerns, with false hierarchy and preselection being the most recurrent mechanisms identified. In the remaining scenarios, classifications were less stable, indicating that some cases require further adjudication. The study indicates that prompt wording can affect not only what prototypes do, but also how they organize choices, defaults, and visual emphasis. This suggests that prompts operate as requirements-like design artifacts during AI assisted interface prototyping, since they can propagate design intent into functional interface structures.