Al could soon replace human taste testers, saving manufacturers time and money during the development of new products.

A new study from the University of Illinois Urbana-Champaign explores how ChatGPT can be used in the sensory evaluation of foods, specifically brownies.

Study author Damir Torrico - assistant professor in the Department of Food Science and Human Nutrition, part of the College of Agricultural, Consumer and Environmental Sciences at Illinois - argued that relying on human testers can slow down the process, especially when multiple prototypes need to be evaluated simultaneously.

"Sensory panels require time and careful coordination, and in some cases, certain ingredients will not be food-grade, making them unsuitable for tasting," said Torrico. "That is why large language models such as ChatGPT are being considered for sensory evaluation. It is possible to create models that can replicate certain human responses."

Put to the test

Torrico provided ChatGPT with 15 different  brownie recipes, ranging  standard ingredient  lists to  unusual  combinations including mealworm powder and fish oil, and asked the Al to describe the sensory  characteristics of  each. The results  were then categorised to determine if the Al's responses were positive, negative or neutral.

ChatGPT's responses were overwhelmingly positive, even for recipes that included unusual ingredients. This positivity aligns with the psychological phenomenon scientists call hedonic asymmetry - the idea that people (and apparently Al) tend to describe things that are beneficial to them in more positive terms.

Since food serves a role in keeping us full and giving us energy, humans respond to food in a positive manner. In its attempt to appear human, ChatGPT se'F(1 t splay this behaviour, Torrico explained.

While kinks like hedonic asymmetry still need to be worked out, the study showcased the potential benefits of using Al in food development. ChatGPT could be used as a screening tool, helping food scientists narrow down recipe options before presenting them to a consumer panel - with the additional benefit of saving the industry time and money.

"Using  Al 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 added. "I could see ChatGPT being developed for sensory evaluation to help the industry."

Torrico's next steps are to refine the experiment, training ChatGPT to respond with a vocabulary that is similar to a human descriptive panel.

Meanwhile, more than £7m has been awarded by UK Research and Innovation to select businesses to help boost Al growth.