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

Area of Science:

• Urban Studies
• Human-Computer Interaction
• Artificial Intelligence

Background:

• Traditional urban walkability assessments using computer vision and machine learning lack subjective and emotional dimensions.
• Large language models (LLMs) offer potential for analyzing unstructured data in urban studies.
• Evaluating LLMs' ability to capture human perceptions of walkability is crucial for designing pedestrian-friendly environments.

Purpose of the Study:

• To critically evaluate if large language models (LLMs), specifically GPT-4o, can accurately reflect human perceptions of urban walkability.
• To compare GPT-4o's evaluations of urban scenes with human judgments across key walkability dimensions.

Main Methods:

• Human participants and GPT-4o evaluated street-level images based on walkability dimensions: overall walkability, feasibility, accessibility, safety, comfort, and liveliness.
• Text mining techniques, including keyword frequency, coherence scores, and similarity indices, were used to analyze responses.
• Comparative analysis between GPT-4o's and human participants' evaluations.

Main Results:

• GPT-4o and human participants showed alignment in evaluating overall walkability, feasibility, accessibility, and safety.
• Significant differences were observed in the assessment of comfort and liveliness, with humans exhibiting broader thematic diversity.
• GPT-4o provided more focused and cohesive responses, particularly for comfort and safety, with moderate similarity scores to human judgments.

Conclusions:

• Human input is essential for comprehensive, human-centered walkability evaluations.
• LLMs demonstrate potential but require refinement to fully align with human perceptions in urban walkability studies.
• Future research should focus on enhancing LLMs for nuanced understanding of subjective urban environmental qualities.