An international team of scientists from Taiwan, Malaysia and Indonesia asked ChatGPT to provide nutritional information for 222 foods to understand how the neural network could help with meal planning. Their research published in JAMA Network Open.

In the experiment, the AI ​​was asked to draw a table for line-by-line calculations of the energy (in kilocalories) / carbohydrates (in grams) / lipids (in grams) / protein (in grams) content of the following foods (raw, uncooked). The researchers judged the accuracy of the neural network’s answers based on the assessments of professional nutritionists based on database food composition from the Taiwan Food and Drug Administration.

As a result, the experts did not notice strong differences in professional and neural network estimates of energy, carbohydrates and fat content. However, they were observed in relation to protein – AI significantly exceeded its amount.

The study found that while neural network responses are sometimes inaccurate, they can still be useful and convenient for people who want to use AI as a source of nutritional information about foods. However, the authors of the work emphasized that now hour-bots are not so developed, so they are not yet able to give advice on a nutrition plan and calculate the amount of food that an individual person needs to consume. The responses from these chats can be influenced by many factors, such as the input language, the clarity of the question, and the software environment.

Anna Morozova