Image of the City in the Age of Large Language Models: ChatGPT as a New Urban Observer

Abstract

This research examines how city images of Istanbul’s Historic Peninsula are evolving with artificial intelligence and expanding user-generated content. City image refers to a mental representation of urban structure, and imageability to the ease of forming it. Using Lynch’s five-element framework (paths, edges, districts, nodes, landmarks), the study compares a classical verbal cognitive map based on 32 interviews with built-environment academics (noting possible sample bias) with ChatGPT’s reading of 210 citizen reviews from TripAdvisor and Ekşi Sözlük. A structured, multi-step prompting workflow, guided and checked by humans at each stage, prompts the model (not trained on the corpus) to extract, categorize, and summarize key locations and experiences in these informal narratives. While the model recognizes familiar Lynch elements, it also redistributes their prominence: iconic buildings become less dominant, while comfort, crowding, safety, and spatial atmosphere become more central. The findings suggest that large language models can function as urban observers capable of summarizing collective city images at scale. The method offers a reproducible way to connect interview-based cognitive mapping with platform narratives. It highlights the need for critical, human-led interpretation when applying language models to urban design.