AI Takes Surprising Bite Into Food Development

Artificial intelligence is now venturing into unexpected territory: evaluating the sensory qualities of chocolate brownies. A study from the University of Illinois Urbana-Champaign reveals that large language models like ChatGPT could streamline food product development and potentially reduce costs for manufacturers navigating an increasingly competitive market.

Food scientist Damir Torrico, an assistant professor in the Department of Food Science and Human Nutrition, found that ChatGPT can effectively generate sensory descriptions for hypothetical brownie formulations, including those containing unconventional ingredients like mealworm powder and fish oil.

“Sometimes, relying on human testers can slow down the process, especially when multiple product prototypes need to be evaluated simultaneously,” Torrico explained in conjunction with the study’s publication. The research highlights how AI could serve as a preliminary screening tool before companies invest in expensive human sensory panels.

“ChatGPT was trying to always see the good side of things,” noted Torrico, pointing out that the AI assigned remarkably high quality scores between 8.5-9.5 out of 10 across all formulations.

The experiment involved fifteen different brownie recipes, ranging from standard formulations to those with unusual ingredient substitutions. ChatGPT was prompted to act as an experienced taster and provide detailed sensory descriptions for each formulation. Researchers then analyzed these responses using Natural Language Processing techniques to identify patterns in sentiment and descriptor usage.

Interestingly, the AI displayed what researchers call “hedonic asymmetry” – a tendency to provide overwhelmingly positive evaluations even for formulations that would likely receive negative reactions from human consumers. This phenomenon mirrors human psychological patterns where we tend to describe beneficial items more positively.

“ChatGPT was trying to always see the good side of things,” noted Torrico, pointing out that the AI assigned remarkably high quality scores between 8.5-9.5 out of 10 across all formulations.

This positivity bias represents both a limitation and an opportunity. While it suggests AI isn’t yet capable of accurately predicting consumer rejection of unusual ingredients, it demonstrates that language models can generate consistent sensory descriptions that might help food scientists narrow their focus during development.

The financial implications for food manufacturers are significant. Traditional sensory evaluation requires recruiting and training panels of human testers, which becomes increasingly expensive as the number of product variations increases. By using AI as a preliminary screening tool, companies could potentially test hundreds of virtual formulations before advancing only the most promising candidates to human testing.

This approach could prove particularly valuable for startups and smaller companies developing novel food products. Reducing the number of physical prototypes needed before market testing could dramatically cut R&D costs and accelerate time-to-market in the competitive food space.

For investors tracking foodtech innovations, this research indicates a potential new direction for AI applications in product development. While most attention has focused on generative AI for marketing and logistics, its application in product formulation and sensory science opens new efficiency pathways.

Despite ChatGPT’s inability to physically taste food, it produced consistent sensory descriptions based on ingredient lists alone.

Sentiment analysis of the AI’s responses showed specific patterns. “In general, the terms ‘trust’, ‘anticipation’, and ‘joy’ were the most frequently expressed sentiments found in the ChatGPT responses. On the other hand, ‘disgust’, ‘fear’, and ‘trust’ were the least frequently expressed sentiments in these responses,” according to the study. Further analysis showed that the standard brownie formulations were associated with descriptors like “texture” and “slight,” while common ingredient replacement formulations were linked to terms such as “chocolate,” “fudgy,” and “flavor.”

The study highlights both current limitations and future potential. Despite ChatGPT’s inability to physically taste food, it produced consistent sensory descriptions based on ingredient lists alone. This suggests that with further refinement and training on specialized datasets, AI could become increasingly valuable in food science applications.

“Using AI can give general insights of what products can be considered for further testing, and what products shouldn’t be put through that long process,” Torrico said. “I could see ChatGPT being developed for sensory evaluation to help the industry.”

Next steps in this research will involve comparing AI-generated sensory descriptions with those from human panels to validate accuracy and develop more sophisticated predictive models. Torrico plans to refine the experiment by training ChatGPT to respond with vocabulary similar to a human descriptive panel.

As competition in alternative proteins and novel food ingredients intensifies, tools that can rapidly evaluate multiple formulations without physical production could become essential competitive advantages. This study suggests that while robots won’t replace human taste testers anytime soon, they might help ensure only the most promising formulations ever reach human palates.

The research was published in the journal Foods and adds to growing evidence that AI applications in the food industry extend well beyond supply chain optimization and consumer marketing.