Can You Rely on ChatGPT to Track Your Nutrition? What the Studies Found

You can use ChatGPT to estimate what is in a meal, but the published tests say not to rely on it alone. Across three studies it recognized foods well, got calories within 10% for fewer than half of the foods it was asked about, and misjudged portions as meals got bigger. It also keeps no food diary of its own: no daily totals you can chart, no targets that adjust, no weight trend. The useful question is not whether to use ChatGPT for nutrition, but what to pair it with.

Calories within 10%

39–47% of 222 foods, ChatGPT-3.5 and 4

Foods recognized in photos

93% of 114 meals’ foods

Nutrients off by more than 10%

13 of 16 in the photo study

PlateLens, measured on plates

1.1% average calorie error across 180 weighed meals, Dietary Assessment Initiative

What the studies found

Calories for single foods. Researchers in Taiwan asked ChatGPT-3.5 and ChatGPT-4 for the calories and macros of 222 foods from eight menus and compared the answers with a nutritionist’s values. Calories came within 10% for 39% to 47% of the foods, depending on the version and the language of the question. The answers were consistent when the same question was repeated, and ChatGPT-4 overestimated protein (Hoang et al., JAMA Network Open, 2023).

A wider tolerance. A second study compared ChatGPT’s values with USDA data and found that 97% of its calorie figures fell within 40% of the reference; the meals in the daily plans it wrote fell within 30% of their USDA calorie values. The authors call that reasonably accurate for planning (Haman et al., Nutrition, 2024). For a food log it is a wide band: 40% of a 600-calorie dinner is 240 calories, about half of a typical daily deficit.

Meals from photos. The third study uploaded 114 meal photographs from a national dietary survey. ChatGPT-4 identified the foods with 93% precision and estimated the weight of small meals well, but its weights for medium and large meals disagreed with the real ones. It was more than 10% off for 13 of 16 nutrients and came in low on 11 of them (O’Hara et al., Nutrients, 2025).

All three tested earlier versions of ChatGPT. A newer model is not automatically a measured one, so these are the best published evidence rather than the last word.

Where ChatGPT’s numbers drift

• The portion. The photo study found the gap exactly where most plates sit: medium and large servings. Naming the food correctly and misjudging how much of it there is gives a confident, wrong total.
• The fat nobody mentions. Oil in the pan, butter on the vegetables and dressing on the salad are rarely visible in a photo or written in a description, and they are the entries that most often go missing, as covered in why a calorie deficit stalls.
• The description. “A bowl of pasta” can hold one serving or three. The estimate is only as specific as what you type.
• The record. A chat thread is not a diary. The totals live in a conversation, not in a record you can correct, chart or average by week, and nothing connects them to your weight trend.

What a verified database fixes, and what it does not

The usual argument against ChatGPT, made by Cronometer among others, is that it has no verified food database. That is fair, and a curated database, such as Cronometer’s, makes each entry right. The photo study shows where the rest of the error lives, though: ChatGPT named 93% of the foods correctly and still missed on 13 of 16 nutrients, because the error was in how much, not what. A verified entry logged with the wrong portion, or without the oil, carries the same error. A food diary is as accurate as the portion and the fat that go into it.

How to use ChatGPT for nutrition without relying on it alone

1. Keep the log in an app that records every meal with its portion and cooking fat, and keeps daily totals, targets and your weight trend.
2. Use ChatGPT for what it does well: questions, meal ideas, higher-protein swaps, explaining a label. Its answers are consistent, which suits explanations better than exact counts.
3. Treat any number it gives you for a meal as a first estimate and log the portion you actually ate.
4. Connect the two, so ChatGPT answers from your real diary instead of a one-line description. How to count calories with ChatGPT walks through the setup.

Where PlateLens fits

The studies above missed in the same places: the portion and the fat nobody mentions. PlateLens logs the plate itself from a photo or a sentence such as “salmon, a cup of rice and a tablespoon of olive oil”, and a correction in plain words, such as “the rice was two cups”, updates the entry. The PlateLens plugin for ChatGPT then gives ChatGPT read-only access to that diary, with your meals, targets, weight trend and energy balance, so its answers come from what you logged. The estimates behind that diary were measured by others: the Dietary Assessment Initiative’s published validation study found an average calorie error of 1.1% across 180 weighed meals, and the open-source Foodvision Bench reproduced it on 231 more on its public GitHub leaderboard.

Frequently Asked Questions

Where this comes from

• Hoang YN, Chen YL, Ho DKN et al. Consistency and accuracy of artificial intelligence for providing nutritional information. JAMA Network Open, 2023.
• Haman M, Školník M, Lošťák M. AI dietician: unveiling the accuracy of ChatGPT’s nutritional estimations. Nutrition, 2024.
• O’Hara C et al. An evaluation of ChatGPT for nutrient content estimation from meal photographs. Nutrients, 2025.
• Independent measurement of PlateLens: the Dietary Assessment Initiative six-app validation study (180 weighed meals) and the Foodvision Bench mini-231 leaderboard.
• The PlateLens plugin listing in ChatGPT’s Plugins directory; setup and permissions: Use PlateLens in ChatGPT.