Why AI chatbots can make up nutrition numbers
By Lance Hughes, founder of Plategist · Updated 3 November 2026
A chatbot writes the most likely-sounding answer, so it can give a nutrient figure that no database holds.
What a 2023 study found
Researchers asked two chatbot models, ChatGPT-3.5 and GPT-4, about 222 foods, five times each, in English and in Chinese, and checked the answers against Taiwan's food composition database (Hoang et al., JAMA Network Open, 2023, doi:10.1001/jamanetworkopen.2023.50367).
- Energy came within 10% of the database figure for 39% to 47% of the foods, depending on the model and the language. So for more than half, the figure was further off than that.
- GPT-4 overestimated protein, the one nutrient that differed significantly from the nutritionists' figures.
The authors still called AI "a useful and convenient tool", while saying it can't replace a nutritionist. And these were 2023 models. The point isn't that chatbots are useless. It's that a generated figure can't be traced back to a row in any database, so there's no way to check where it came from.
How Plategist splits the job
In Plategist, the AI does two things: it names the food and guesses the grams. It doesn't write the numbers. Every number comes from AUSNUT, and the database row it came from is named on every food, so you can check it.
That doesn't make the estimate exact. It makes it checkable.
Plategist estimates what's on your plate from a photo, then looks each food up in AUSNUT. 14 days free, then a monthly or yearly plan.