LLMs Unreliable for Cosmetic Chemistry and Skincare Advice

With consumers increasingly seeking skincare advice from AI chatbots, a critical evaluation of Large Language Models (LLMs) in cosmetic chemistry and skin health was conducted. A study benchmarked 14 LLMs on structured topics, including chemical properties of ingredients and common skincare scenarios, with web search disabled to assess internalized knowledge. The overall performance was found to be poor, with significant weaknesses in quantitative reasoning and identifying chemical structures. Although LLMs could provide generally reasonable answers to broad skincare questions, their responses consistently lacked the technical depth necessary for consumers to make truly informed decisions. A notable risk identified was that authoritative-sounding but technically incorrect outputs were less likely to prompt skepticism than responses acknowledging uncertainty. The findings strongly suggest that general-purpose LLMs, primarily trained on unverified public data, are currently not dependable sources for cosmetic chemistry information. Future improvements will likely require fine-tuning on verified chemical and dermatological datasets, alongside substantial advancements in algorithmic reasoning.