Comparing Artificial Intelligence (AI) generated sensory responses with human descriptive analysis using chocolate brownies

Artificial Intelligence (AI) has been increasingly discussed as a potential mechanism to support traditional sensory evaluation, which is limited by high cost, time, and human bias. This study evaluated non-traditional flavor ingredients in chocolate brownies as a food model system to assess the alignment of AI-generated responses to human sensory descriptive analysis and physicochemical properties. Ten chocolate brownie formulations were developed using conventional and non-traditional ingredients (stevia, gochujang, barbecue, miso, chicken stock, fish sauce, garlic, liquid smoke, Splenda, and control/classic), designed with modified attributes in taste, aroma, and chemical sensations. Brownies were prepared and analyzed for color, moisture content, and texture (texture profile analysis). Human sensory data were collected from a trained panel group (n=20) using a hybrid of Spectrum and Quantitative Descriptive Analysis (QDA) approach with a 15-point intensity scale for sensory attributes. Each formulation was evaluated in duplicates. AI sensory responses were generated using ChatGPT 5.4 mini, in which 20 virtual panelists were prompted and acted as a trained internal panel to evaluate the ten formulations. Human sensory data, AI-generated data, and instrumental data were analyzed using mixed-model ANOVA, radar plot, and principal component analysis (PCA). Results showed significant differences in color (crumb redness and crumb yellowness), hardness, and moisture content across formulations. Regarding sensory outputs, ChatGPT generally overestimated responses to basic tastes and aromas. However, ChatGPT was closely aligned with human responses in terms of texture descriptors. In savory formulations (miso and fish oil), ChatGPT had a positive bias towards “bitterness” and “umami”. On the other hand, marginal differences were found for “density” and “chewiness”. This study shows that AI tools such as ChatGPT can be potentially used to screen different formulations. Further studies across a wide range of food categories are suggested to integrate stronger AI into sensory evaluation and support sensory-driven product development.