CHAMPAIGN, ILLINOIS — Artificial intelligence continues to shape diverse industries, and food science is no exception. A new study from the University of Illinois Urbana-Champaign has spotlighted the growing potential of ChatGPT in sensory evaluation of foods, focusing specifically on brownies. The findings point toward the possible integration of AI in streamlining new product development and assisting in early-stage recipe analysis, potentially saving time and resources traditionally required for human sensory panels.
The research team, led by Damir Torrico, assistant professor in the Department of Food Science and Human Nutrition, conducted a detailed examination involving 15 unique brownie recipes. These ranged from conventional chocolate-and-butter combinations to more unconventional formulations incorporating ingredients like mealworm powder and fish oil.
Torrico fed these diverse recipes into ChatGPT, prompting it to describe each brownie’s sensory characteristics—taste, texture, mouthfeel, and overall appeal. The AI-generated descriptions were then assessed and classified as either positive, negative, or neutral.
What emerged was a pattern of generally positive feedback from ChatGPT, regardless of the inclusion of controversial ingredients. Brownies made with insect proteins or marine-based fats still received largely favorable descriptions. According to Torrico, this echoes a phenomenon in psychology called “hedonic asymmetry.” The term refers to the innate human tendency to describe beneficial or pleasurable experiences using favorable language. In replicating human behavior, ChatGPT appeared to mirror this bias, consistently highlighting the positive traits of food, even in unconventional formats.
This trend opens up compelling questions about the potential role of large language models (LLMs) like ChatGPT in replacing or augmenting traditional sensory panels, especially during the initial phases of product development.
Bridging AI and Human Palates
The motivation behind the study stemmed from the logistical challenges often associated with human sensory testing. Traditional sensory panels involve significant time investment, scheduling, and cost, not to mention the limitations imposed by the need for food-grade ingredients when testing is conducted on humans. Some early-stage prototypes may use non-food-safe components or contain unfamiliar additives that aren’t fit for immediate consumption.
By contrast, AI models can process multiple formulations simultaneously and generate preliminary feedback within seconds. This ability to provide instant qualitative assessments makes them an attractive tool for food scientists and manufacturers aiming to iterate rapidly on product design before investing in more detailed human testing.
“AI doesn’t replace the human experience,” Torrico clarified. “But it can definitely serve as a preliminary screening tool to identify which formulations are promising and which may not warrant further exploration.”
Understanding the Bias in AI Taste Perception
Interestingly, the uniformly positive tone of ChatGPT’s evaluations may be a double-edged sword. While its encouraging remarks may streamline product development, they also risk overlooking or underplaying potential flaws in a recipe. If an AI consistently avoids negative feedback, it could inadvertently skew product assessments toward optimism.
This optimistic tone, however, is not entirely unexplainable. Torrico believes that AI’s attempt to emulate human behavior may naturally favor a positive interpretation of food-related stimuli. Since food is fundamentally associated with sustenance and pleasure, the language used to describe it—even by an AI trained on human text—tends to be favorable.
The challenge, therefore, lies in refining these models so that they can provide balanced, nuanced descriptions, similar to how trained sensory panelists might respond. Torrico has plans to explore this dimension further by custom-training ChatGPT to adopt a more descriptive, technical vocabulary, commonly used in formal sensory evaluation. This could include using standardized terms such as “moist crumb,” “astringency,” “bitterness,” or “aftertaste” to enhance the objectivity and detail of its evaluations.
Potential Industrial Applications
The food industry is no stranger to leveraging technology for innovation, from automated production lines to digital inventory management. The inclusion of AI in sensory evaluation introduces another dimension of efficiency, particularly in areas such as:
- Early-stage R&D: ChatGPT could help product developers screen dozens of recipe variations quickly, reducing the workload on human testers.
- Concept validation: For novel ingredients such as alternative proteins, ChatGPT could simulate human response and predict consumer acceptance trends.
- Market forecasting: With refined training, AI could eventually be used to assess how certain flavor profiles or textures might perform in specific markets or demographics.
Given these possibilities, Torrico’s study represents a foundational step toward integrating AI-driven tools in food science workflows. While ChatGPT is not yet ready to replace trained sensory panels, its use as a cost-effective, scalable companion in early evaluations could soon become standard practice.
The Path Ahead: Training AI to Taste Like a Human
The researchers acknowledge that the current version of ChatGPT, while powerful, was not initially designed for domain-specific sensory tasks. As such, the next phase of the study will focus on refining the AI’s lexicon and contextual understanding of sensory language. This could involve feeding it databases of professional tasting notes, technical sensory reports, and terminologies used by food scientists and culinary professionals.
Once trained in this specialized vocabulary, ChatGPT could potentially mimic not just the sentiment of human reactions, but also the depth and precision of their observations. This would make it more viable for high-stakes product development, such as health-focused formulations or premium food items aimed at niche markets.
Moreover, such a tool could be integrated into existing product development software, giving food companies a competitive edge in accelerating innovation without compromising on sensory quality.
Challenges in AI Adoption
Despite the promise, the integration of ChatGPT into food evaluation does come with caveats. The reliability and consistency of AI-generated sensory analysis need to be tested across a broader array of food categories beyond brownies. Items with more complex flavor profiles, such as fermented products, alcoholic beverages, or regional ethnic cuisines, may present a tougher challenge for AI to evaluate meaningfully.
Also, regulatory frameworks in food safety and labeling may eventually need to account for AI’s role in formulation and assessment. As more companies consider AI tools in product development, ethical and accuracy concerns are likely to arise, particularly when AI is involved in describing items for public consumption or nutritional claims.
Still, the potential for AI to become an indispensable tool in the food industry remains strong. The findings from Torrico’s study contribute a fresh perspective to the broader conversation on how AI can enhance human expertise, rather than replace it.
The University of Illinois study presents an insightful case for how ChatGPT can be harnessed for preliminary sensory evaluation of food products, specifically brownies. Its ability to generate quick, descriptive feedback holds promise for accelerating product development in the food sector. While limitations exist—such as a bias toward positive descriptions and a lack of technical vocabulary—the research lays the groundwork for future integration of AI-assisted tasting simulations. With further refinement, tools like ChatGPT could help bridge the gap between early product concepts and consumer-ready offerings, streamlining innovation in a traditionally labor-intensive process.
