Place identity: a generative AI's perspective

The concept of place identity, the unique characteristics and meanings associated with a location, is central to environmental psychology, urban design, and geography. Traditional methods for studying place identity, such as qualitative interviews and photo-elicitation, are time-consuming and may yield biased results due to small sample sizes. The advent of user-generated content and, more recently, generative AI offers new avenues for understanding place identity at scale. This study investigates the potential of GenAI, specifically ChatGPT (for text) and DALL-E2 (for images), to generate realistic representations of place identity for 64 global cities. The researchers address two key questions: 1) How does GenAI illustrate place identity? and 2) To what extent can we trust GenAI's representation of place identity when compared with fact-based data?

Prior research on place identity highlights the complex interplay between physical settings, events, and individual/group meanings in shaping distinctive place identities. The distinction between 'space' (lacking meaning) and 'place' ('center of felt value') is crucial. Understanding place identity aids urban planning by facilitating the creation of livable and legible places fostering community attachment and sustainable practices. However, measuring place identity is challenging due to its subjective nature. While existing studies used qualitative methods or leveraged user-generated content (text and images) with NLP and computer vision, this study explores the novel application of GenAI models.

The study employed a computational framework with two primary steps: (1) exploring place identity with GenAI and (2) validating results against real-world settings. For text, the researchers prompted ChatGPT with variations of the question, "What is the place identity of [city]?". For images, DALL-E2 was prompted with "What is the place identity of streetscapes in [city]?" resulting in 20 images per city. For validation, textual data from Wikipedia introductions and image data from Google image searches were collected. Text similarity was assessed using a sentence transformer BERT model (calculating cosine similarity between sentences from ChatGPT and Wikipedia). Image similarity was measured using Learned Perceptual Image Patch Similarity (LPIPS), calculated using AlexNet as the feature extractor, generating image similarity scores (S = 1 – LPIPS). A human survey (30 participants) rated the similarity between DALL-E2 and Google images on a 7-point Likert scale. Finally, city-by-city visual similarity was assessed using normalized Chamfer distance (CD) between DALL-E2 image sets.

ChatGPT successfully captured various aspects of place identity, including economic, educational, cultural, and historical elements, often incorporating proper nouns that signify unique place characteristics. The average text similarity scores between ChatGPT and Wikipedia were 0.59 (place), 0.58 (urban), and 0.56 (street), suggesting that the AI could capture important aspects of place identity but also highlighting discrepancies that may be related to the length of generated texts. Analysis of word clouds further showed the overlap in key themes between ChatGPT and Wikipedia. Concerning images, DALL-E2 produced visually recognizable representations of place identity for some cities but generated generic city views for others, underscoring the challenges of capturing nuanced place-specific attributes in visual form. The average image similarity (S) score was 0.575. A positive but weak correlation (0.229, significant at the 0.1 level) was found between LPIPS-based image similarity and human ratings. The pairwise comparison using normalized Chamfer distance revealed higher similarity scores within the same city than across different cities. Cities in the Americas and Europe demonstrated higher visual similarity among themselves compared to cities in other regions. This suggests that GenAI models may capture both place-specific characteristics and broader regional trends.

The findings demonstrate that GenAI models possess the potential to generate valuable data on place identity. The ability of ChatGPT to generate relevant text descriptions and the ability of DALL-E2 to create visually distinctive images for some cities are promising. However, the inconsistencies and lower similarity scores for some cities suggest limitations in the accuracy and generalizability of GenAI models, particularly highlighting challenges in capturing the subjective nuances of place identity. The study's findings also hint at potential biases, especially concerning visual representations of places outside the Western world, suggesting a need for more comprehensive and representative training datasets. The weak correlation between human judgments and LPIPS scores points to the importance of incorporating human-in-the-loop evaluations when using GenAI models for subjective concepts like place identity.

This study offers a pioneering investigation into the use of GenAI for understanding place identity in urban studies. While GenAI shows promise in capturing salient city characteristics, its limitations concerning accuracy, biases, and the subjective nature of place identity must be acknowledged. Future research should focus on improving prompt engineering, refining similarity assessment methods, incorporating multi-source data fusion, and employing explainable AI to enhance the reliability and transparency of GenAI models in representing place identity. Further research should also investigate cultural and linguistic variations in GenAI's ability to capture place-specific meanings.

The study's findings are limited by the specific GenAI models (ChatGPT and DALL-E2) and datasets used. The selection of 64 cities may not fully capture the diversity of global urban contexts. The reliance on Wikipedia and Google Images for ground truth introduces potential biases and limitations, as these sources themselves may not always reflect a perfectly accurate or complete representation of a city's place identity. The relatively small sample size in the human survey may also limit the generalizability of findings concerning human perception of image similarity. The English-only approach for both text and image generation introduces bias and limits the study's applicability to other cultures and languages.