Examining Linguistic Dynamics and Inclusion inText-to-Image Generative AI for Participatory Architecture

Sammanfattning

This study addresses two interrelated objectives: to evaluate how text-to-image generative artificial intelligence reduces technical, engagement, and visual barriers in participatory architecture, and to investigate how this technology influences participants’ narratives when envisioning urban spaces.One-on-one, semi-structured interviews were conducted with 21 participants. These interviews aimed to obtain spatial narratives related to a proposed public park with an integrated Community hub. Following each interview, participants independently translated their narratives into text prompts and generated corresponding visualizations using the text-to-image generator Adobe Firefly. A convergent mixed-methods analysis was applied to the interview transcripts and prompt texts, measuring lexical diversity, Jaccard Index similarity, and the levels of emotional and spatial engagement reported by participants when describing the urban space. This analysis was complemented by a review of the generated images and participant feedback on their experience using the text-to-image generator.Data gathered from the feedback interviews revealed that a majority of participants expressed positive attitudes toward using the text-to-image generator. They perceived it as a valuable tool that significantly enhanced their ability to visualize and refine design and visual concepts. However, content analysis revealed a shift in narrative style. Ideas expressed in prompts tended to become more object-oriented and less socially or emotionally grounded than the interview narratives. This shift suggests that in prompting their ideas, participants reduced emotional, cultural, and social nuance, inadvertently transforming their urban visions. Additionally, despite the diversity of prompts, the images generated frequently converged on Westernized aesthetics. This outcome aligns with concerns raised in prior literature about aesthetic and cultural bias in text-to-image generative artificial intelligence, and it underscores the risk of narrowing the expressive and inclusive potential of participatory architecture.If architects, designers, and planners wish to effectively leverage this technology in participatory architecture processes, it is essential to understand both its limitations and potential. Thoughtful and aware application is crucial to ensure that text-to-image generators support, rather than constrain, diverse and meaningful participations.