Urban walkability through different lenses: A comparative study of GPT-4o and human perceptions

Wedyan, Musab; Yeh, Yu-Chen; Saeidi-Rizi, Fatemeh; Peng, Tai-Quan; Chang, Chun-Yen · 2025 · OpenAlex-citations

DOI: 10.1371/journal.pone.0322078

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study investigates whether large language models (LLMs), specifically GPT-4o, can accurately replicate human perceptions of urban walkability. While traditional computer vision and machine learning methods effectively quantify physical attributes of urban environments, they often fail to capture subjective, emotional, and contextual dimensions. The authors address this gap by comparing GPT-4o’s multimodal evaluations of street-level images against human judgments across six key dimensions: overall walkability, feasibility, accessibility, safety, comfort, and liveliness. The research aims to determine if LLMs can serve as reliable proxies for human-centered environmental assessment. The methodology involved a comparative analysis using 48 pairs of urban street images collected from Michigan, USA. Human participants ($N=174$) evaluated these image pairs via an online survey, rating each dimension on a scale of 1–10 and providing textual justifications. GPT-4o was prompted with the same image pairs using self-consistency techniques, generating 15 responses per pair to ensure reliability. The study employed text mining, including keyword frequency analysis, coherence scoring, and cosine similarity indices, to compare the reasoning and sentiment of human and AI responses. Statistical tests, including independent samples t-tests, were used to evaluate rating differences. Results indicated that GPT-4o and human participants aligned closely in their evaluations of overall walkability, feasibility, accessibility, and safety. Statistical analysis showed no significant difference in mean ratings between humans and GPT-4o for either image set. However, notable divergences emerged in the assessment of comfort and liveliness. Human participants favored the first image in comfort assessments (62%) more than GPT-4o did (47%), while GPT-4o leaned toward the second image. Textual analysis revealed that human responses exhibited broader thematic diversity and addressed a wider range of topics, whereas GPT-4o produced more focused and cohesive responses. Cosine similarity scores averaged approximately 0.46, indicating a moderate level of alignment in reasoning but highlighting distinct differences in how subjective qualities are articulated. The study concludes that while GPT-4o can approximate human judgments for objective or structural aspects of walkability, it struggles to fully capture subjective experiences like comfort and liveliness. Human input remains essential for comprehensive, human-centered urban evaluations. The findings underscore the need to refine LLMs to better align with human perceptions, particularly regarding emotional and contextual nuances. This research provides a critical benchmark for the emerging application of generative AI in urban planning, suggesting that LLMs should currently be viewed as complementary tools rather than replacements for human assessment.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• pedestrian behavior perception