Generative artificial intelligence perspectives on typical landscape types: Can ChatGPT compete with human insight?

Abstract

The emergence of ChatGPT, a prominent generative artificial intelligence (GAI), has raised concerns due to its increasing capability to rival or even surpass human performance across various tasks and domains. However, its alignment with human perception, particularly in emotional and aesthetic dimensions such as landscape preferences, remains uncertain. This study investigated the discrepancies between human and GPT-4 performance in landscape perception and preference, using the Kaplans’ preference matrix as a benchmark. Survey data were collected from 1,333 participants in China, and five typical landscapes i.e. gray, open green, partly open/closed green, closed green, and blue spaces were evaluated. To simulate human-like responses, artificial intelligence (AI) agents using ChatGPT were created with personal attributes mirroring those of the human sample. Results indicated that GPT-4 demonstrated significant divergences from human perception and preference in assessing landscape coherence, complexity, mystery, legibility, and overall preference. While GPT-4 performed comparably well in simpler environments, such as pure single-layer broadleaf forests on flat terrain, it struggled to capture key elements and emotions in more complex or nuanced urban landscapes. Notably, only 2.4 % of ChatGPT’s responses aligned with human perceptions and preferences. These findings highlighted the limitations of current AI in fully replicating human intelligence in landscape perception, emphasizing the continued necessity of human involvement in human-centered landscape design. This study offers insights into the current limitations of ChatGPT and suggests directions for enhancing its application in landscape design.