Research note: assessing human preferences for natural landscapes—an analysis of ChatGPT-4 and LLaVA models

Abstract

This study explores the potential of large language models (LLMs) to approximate human preferences for and aesthetic judgments of natural landscapes using natural language processing techniques. Our research addresses the gap in understanding how well LLMs can replicate complex human perceptions related to landscape preferences. We compared human responses and model predictions across 30 natural scenes in five landscape preference dimensions—complexity, coherence, legibility, mystery, and overall preference. Responses from 50 human participants formed the benchmark for assessing predictions by Chat Generative Pre-Trained Transformer (GPT)-4 and Large Language and Vision Assistant (LLaVA). Correlations between human responses and model predictions evaluated the extent of AI’s ability to mimic complex human perceptions. The results indicate that GPT-4 and LLaVA align significantly with human judgments of complexity, coherence, mystery, and overall preference but not of legibility, which highlights the challenge of evaluating nuanced aspects of natural landscapes using LLMs.