Can you trust nutrition numbers from ChatGPT or Claude?

Updated October 27, 2026

What studies found

• The newest chatbots still missed exact diet targets. In February 2026, ChatGPT-5.2, Claude 4.5 and Gemini 3 wrote 48 one-day diet plans for people with phenylketonuria, who need an exact daily amount of one amino acid. Recalculated with a nutrition database, 36 of the 48 plans (75%) gave less than the prescribed amount. Nutritional Accuracy and Clinical Applicability of Large Language Model-Generated Therapeutic Diet Plans for Phenylketonuria: A Comparative Simulation Study. Nutrients, 2026.
• Calories from a meal photo were off by 36% on average. ChatGPT-4o and Claude 3.5 Sonnet estimated the weight and calories of 52 weighed meals from photos. Their mean error was 35.8% for calories and 36% to 37% for weight. Gemini 1.5 Pro was off by 64% for calories. The authors call the two better models comparable with people recording their own meals. Fridolfsson J, Sjöberg E, Thiwång M, Pettersson S. Performance Evaluation of 3 Large Language Models for Nutritional Content Estimation from Food Images. Current Developments in Nutrition, 2025.
• Calories for single foods were within 10% less than half the time. GPT-3.5 and GPT-4 estimated the calories of 222 foods. Compared with a food composition database, both were within 10% for only about 35% to 48% of the foods. These were 2023 models. Hoang YN, Chen YL, Ho DKN, et al. Consistency and Accuracy of Artificial Intelligence for Providing Nutritional Information. JAMA Network Open, 2023.
• Diets for people with food allergies could be harmful. Researchers asked ChatGPT for 56 diets for people with food allergies. They found it “generally accurate” but with “the potential to produce harmful diets”. The most common errors were in portions and calories. Niszczota P, Rybicka I. The credibility of dietary advice formulated by ChatGPT: Robo-diets for people with food allergies. Nutrition, 2023.

Why it happens

• It does not look the food up. A language model writes the most likely answer from patterns in its training data. It does not read a nutrition database unless one is connected.
• It guesses the portion. “A bowl of pasta” or “a chicken thigh” has no fixed weight. The model picks one and does not tell you.
• It does arithmetic in words. Totals and percentages are written, not calculated, so a sum can be off.
• It can answer differently each time. Ask the same question twice and the numbers can change.
• It writes recipes on the spot. Quantities, cooking times and substitutions are not checked against a source, and a recipe can break the diet you asked for.

Where chatbots help

Chatbots are good at ideas, explanations and planning. They can explain what fiber does, suggest dinners you have not tried, and turn a plan into a list. In one large benchmark, their estimates were comparable to professional nutritionists’ and much faster (NutriBench). The problem is the numbers and the details, which need a source.

How Avocart computes its numbers

In each study above, the correct values came from a nutrition database. Avocart computes its numbers the same way, from the USDA database and the recipe’s own amounts.

1. Avocart reads each ingredient line for its amount and food, such as “2 lb bone-in chicken thighs”.
2. It converts the amount to grams and removes what you do not eat, such as bones, shells and peels.
3. It takes the food’s nutrients per 100 g from USDA FoodData Central and adds up the recipe.
4. It adjusts for cooking where the evidence is clear, such as alcohol in a braise and marinades you discard.
5. It divides by the servings and checks the result. Protein, carbs and fat must match the calories within 15%. Recipes that fail a check are not shown.

The result is calories, protein, carbohydrates, sugar, fat, saturated fat, fiber, sodium and 10 vitamins and minerals per serving. A version of a recipe for another diet gets its own numbers from its own ingredients. The same recipe gives the same numbers every time.

These are still estimates. Brands, portions and cooking change the result. If you have an allergy, always check the product label. Avocart does not provide medical advice.

Get computed numbers in your assistant

Connect ChatGPT or Claude to Avocart and your assistant quotes Avocart’s numbers instead of writing its own. It also sees each person’s diet, targets and today’s activity, and it can log what you eat and fill an Instacart cart.

Common questions