In the rapidly evolving landscape of artificial intelligence, the ability of AI models to accurately represent spatial and cultural identities through imagery is not just a technical challenge but a societal concern. Recent research spearheaded by Virginia Tech’s Junghwan Kim, a noted geospatial data scientist in the College of Natural Resources and Environment, delves into the nuanced performance of AI-generated images in accurately depicting cities of varying sizes. The study meticulously evaluates the efficacy of OpenAI’s DALL·E 2 in visualizing the unique characteristics of small towns versus large metropolitan areas, exposing a significant disparity that could influence public perception and digital representation.
Kim’s initial encounter with AI image generation involved Blacksburg, Virginia—a small university town known for its distinct local architecture and community feel. When prompted to create images of Blacksburg, the AI-generated visuals appeared generic, lacking the unmistakable elements that define the town’s identity. Landmarks that locals hold dear, such as the Hokie Stone-clad buildings of Virginia Tech, were conspicuously absent, raising questions about the algorithm’s ability to capture local essence beyond superficial cityscapes.
Contrastingly, when the same AI system generated images of larger cities like Richmond, Virginia Beach, and the nation’s capital, Washington, D.C., the results showed markedly better alignment with the recognizable and culturally significant features of these urban centers. These images depicted iconic landmarks, waterfronts, and district layouts that contribute to the cities’ identities, suggesting that the AI’s training data and pattern recognition are heavily skewed towards larger locales with extensive photographic and digital footprints.
This discrepancy prompted Kim and a collaborative team from Hong Kong University of Science and Technology (Guangzhou) and the University of Alabama to investigate whether generative AI tools inherently favor larger metropolitan areas over smaller towns. Their findings, published in the peer-reviewed journal Technology in Society, systematically assess the realism and cultural fidelity of AI-generated cityscapes through structured resident evaluations, highlighting a bias that reflects underlying data availability imbalances.
The investigation involved a survey with 129 participants familiar with the target cities. Participants assessed AI-generated images based on their visual authenticity and the extent to which the images encapsulated each city’s unique identity. The evaluation framework was grounded in urban design principles derived from Kevin Lynch’s seminal work, emphasizing landmarks, districts, paths, edges, and nodes. The outcome revealed that images of larger cities consistently achieved higher scores in both realism and accurate identity portrayal.
One of the study’s critical insights was the differential detection of inaccuracies based on respondents’ familiarity with the cities. Long-term residents, possessing deeper local knowledge, were significantly more critical of AI-generated images, noting missing or misplaced features and a dilution of the cultural narrative. This specificity underscores the limitations of generative models that rely mostly on generalized, image-heavy datasets and lack intimate understanding of localized urban fabric.
The broader implication of this research touches on the underrepresentation of smaller communities in digital media ecosystems—a phenomenon that AI models inadvertently amplify. Since these systems learn from massive amounts of internet data, areas with fewer online images and limited digital documentation are prone to being sidelined or misrepresented. Consequently, AI-generated imagery risks perpetuating a digital divide where larger cities dominate the visual and cultural imagination, while smaller towns are rendered invisible or homogenized.
Kim stresses that the findings highlight an urgent need to cultivate more geographically diverse and inclusive datasets during AI model training processes. Integrating local knowledge, ethnographic data, and participatory inputs could enrich the AI’s capacity to generate authentic images that resonate with smaller communities’ identities. Without such interventions, AI’s role in urban visualization and planning could reinforce systemic biases and skew perceptions for users reliant on these images in travel, marketing, and public communication.
The ethical dimensions of AI-generated content further emphasize the importance of transparency and critical engagement with these technologies. While generative AI offers unprecedented speed and creative potential to democratize design and visualization, it simultaneously presents new challenges around representation, accuracy, and cultural sensitivity. For planners, designers, and policymakers, understanding where AI excels and where it falters is essential for responsibly leveraging these tools.
Moreover, the study’s contribution to human geography and urban studies lies in its bridging of computational techniques with spatial cognition frameworks. By employing Lynch’s model, the researchers articulate a structured approach to evaluate not just aesthetic quality but the cognitive and experiential authenticity of AI-generated environments. This interdisciplinary methodology paves the way for subsequent work on AI’s role in mapping, storytelling, and place-making.
As AI continues to permeate domains like tourism, smart city initiatives, and digital marketing, the uneven representation identified by Kim and colleagues serves as a cautionary tale. Generative models, often treated as neutral or objective, are inevitably shaped by their training data’s biases and gaps. Addressing these gaps requires concerted efforts from data scientists, urban scholars, and community stakeholders to co-create datasets and design protocols that honor the diversity of human settlements.
In essence, the research underscores a pivotal dialogue about AI’s future in shaping our visual cultural landscapes. It challenges developers and users to scrutinize the narratives these models propagate and to push for inclusivity and authenticity. As Kim eloquently puts it, generative AI is a powerful tool that must be wielded with mindful awareness of who may be excluded or misrepresented in its digital visions.
The study, “Imagining the City: Evaluating Visual Realism and Identity in AI-Generated Cityscapes Using DALL·E 2,” is a landmark contribution that blends technical rigor with cultural sensitivity. It calls upon the scientific community to prioritize comprehensive geographic representation and elevate local expertise in developing AI systems that inform how we see and understand urban spaces in the digital age.
Subject of Research:
Evaluation of AI-generated cityscape images focusing on visual realism and cultural identity, with an emphasis on disparities between representations of large metropolitan areas and smaller towns.
Article Title:
Imagining the City: Evaluating Visual Realism and Identity in AI-Generated Cityscapes Using DALL·E 2
Web References:
https://doi.org/10.1016/j.techsoc.2026.103360
http://dx.doi.org/10.1016/j.techsoc.2026.103360
Image Credits:
Photo by Max Esterhuizen for Virginia Tech.
Keywords
Artificial intelligence, Generative AI, Urban planning, Geography, Computer science, Visual representation, Cityscapes, AI bias, Geospatial data, Urban studies
