Answers are grounded in this paper and cite their sources.
This study examines the use of a large language model (LLM), specifically ChatGPT 3.5, to train novice formulators in the development of personal care products. The aim is to assess the LLM’s ability to guide students as they formulate a 10-minute hydrating face mask. The research explores how effectively students can rely on the LLM for ingredient substitutions and recipe adjustments during an iterative formulation process, with the goal of producing a high-quality or improved product. Results indicate that while ChatGPT 3.5 demonstrates above-average chemistry knowledge and can provide useful suggestions when prompted clearly, it has significant limitations.1 These include unreliable memory in extended conversations and difficulty with precise mathematical calculations, particularly for ingredient adjustments. For example, the LLM’s limited memory hindered its ability to incorporate information from earlier iterations, often resulting in redundant or inconsistent recommendations. To address these calculation errors, the team developed in-house code to ensure formulation accuracy. Additionally, the LLM’s contribution to cost optimization was minimal, and it struggled to identify complex formulation components that trained formulators would typically recognize. Although the LLM supported rapid initial product development, it was less effective in more advanced stages, including cost optimization and refining complex components.
No takes yet. Share an insight, caveat, or question.
Bilbao et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: