Automating chemosensory creativity assessment with large language models

Chemosensory creativity, the ability to innovate using taste and smell, is a crucial yet understudied aspect of human ingenuity. This study explores the potential of large language models (LLMs), specifically GPT-4, to assess creativity in the context of flavour pairings. Leveraging a novel chemosensory creativity test inspired by the Alternative Uses Task, 200 UK-based participants generated flavour pairings across four food categories. Subsequently, these pairings were rated for creativity, deliciousness, and surprise by human participants and two GPT-4 configurations: a deterministic (low-randomness) model termed “Strict GPT” and a stochastic (high-randomness) model termed “Flexible GPT.” Findings reveal a strong correlation between human and Flexible GPT ratings of creativity ( r = 0.89), surpassing that of Strict GPT ( r = 0.71). Both humans and GPT models relied heavily on novelty (operationalised as surprise) rather than functionality (operationalised by deliciousness) as a determinant of creativity. However, GPT ratings exhibited a stronger emphasis on novelty and higher variability. While GPT-4 demonstrated strong potential for approximating human assessments in the UK context, differences emerged, particularly for rare flavour pairings where human ratings and model predictions had less alignment. These findings demonstrate the feasibility of using LLMs to automate creativity assessments in food-related domains by approximating human evaluations. • Humans and large language models (GPT-4) evaluated creativity in food flavour pairings. • Flexible GPT-4 ratings showed strong alignment with human creativity assessments. • Novelty emerged as the key determinant for both human and AI creativity evaluations. • Human and AI creativity judgments showed least agreement for rare flavour combinations. • The study used a UK-based participant sample, which limits the cultural generalisability of the findings.