The credibility of dietary advice formulated by ChatGPT: robo-diets for people with food allergies
Paweł Niszczota Nutrition 11 May 2023, 112076
Highlights
• We applied a prominent large language model (ChatGPT) in nutritional sciences.
• ChatGPT was tested via 56 diets for 14 food allergens and at 4 restrictions levels.
• We tested the safety, accuracy and attractiveness of these ‘robo-diets’.
• ChatGPT produced balanced diets, but it was unsafe for one allergen.
• We discussed how the quality of robo-diets can improve in the future.
The introduction of ChatGPT has sparked enormous public interest in large language (deep-learning) models, which have been sophisticated enough to perform well on a variety of tasks. One way people are utilizing these models is to construct diets. The prompts often include food restrictions which are an obligatory part of everyday life for millions of people around the world. We investigate the safety and accuracy of 56 diets, constructed for hypothetical people that are allergic to food allergens. Four levels, corresponding to the ‘baseline’ ability of ChatGPT without prompting for specifics, as well as its ability to prepare appropriate diets when a person has an adverse food reaction to two allergens or solicits a low-calorie diet, were defined.
Our findings show that ChatGPT – while generally accurate – has the potential to produce harmful diets. More common errors involve inaccuracies in portions or calories of food, meals, or diets. We discuss how the accuracy of large language models could be increased and the tradeoffs involved. We propose that prompting for elimination diets can serve as one way to assess differences between such models.
