How accurate are AI photo calorie counters? What studies show

AI photo calorie counters are typically off by about 20–35% on a single meal in published tests, mostly because of portion size. That is better than most people do by eye, but not the accuracy of a food scale. Used consistently, with a few corrections, a photo estimate is good enough for most tracking goals.

What studies actually measured

• ChatGPT-5 on 195 dishes (Nutrients, 2025). From the image alone, the average calorie error was 35% on recipe-site dishes, 28% on US meal photos and 20% on home-cooked meals that a dietitian had weighed. Adding standard descriptors, and then an ingredient list with amounts, made the estimates better.
• Nutrition5k (Google Research, 2021). A model trained on 5,000 real dishes whose ingredients were weighed predicted calories from a single photo with a 26.1% average error (70.6 kcal). Using depth-sensor data brought it down to 18.8%, but phones do not have that. In the same paper, people estimating the weight of dishes by eye were wrong by 53% on average (non-nutritionists) and 41% (nutritionists).
• 40 vision models (Scientific Reports, 2026). The best AI models still lagged professional nutritionists, photographing a dish from several angles did not help, and a good smartphone photo beat a controlled lab photo.

Self-report is not perfect either. In a New Zealand pilot, both a photo-based app and 24-hour recall estimated intake about 410–430 kcal a day (1,715–1,814 kJ) below measured energy expenditure, and the two methods did not differ significantly from each other. Nobody gets perfect numbers without weighing.

Which meals are hardest?

The ChatGPT-5 study split dishes by size. Small, low-calorie dishes (280 kcal or less) had the highest percentage error, about 55%, but a modest absolute error of about 96 kcal. Dishes above roughly 500 kcal had a lower percentage error, about 31%, but the largest absolute error, about 216 kcal. In practice, a big restaurant plate is where a few minutes of editing portions saves the most calories, while a snack that is a bit off barely moves your day.

Why portions are the hard part

A photo shows what is on the plate, not how much. The main reasons estimates drift:

1. Depth is invisible. A camera cannot easily tell a 200 g bowl of rice from a 300 g one. Cooked white rice is 130 kcal per 100 g, so a one-cup (158 g) serving is about 205 kcal, and a 300 g mound is 390.
2. Hidden fat. Oil, butter, ghee and dressing disappear into food. One tablespoon of olive oil is 119 kcal.
3. Mixed dishes. Stews, curries and layered dishes hide ingredients, so the model has to guess the recipe.
4. Plates and angles. A big plate makes portions look small; a very close or dark photo hides edges.

How to get a better estimate

• Take one clear, top-down photo in good light with the whole plate in frame. Extra angles did not help in the tests.
• Tell the app what you know. Ingredient information improved accuracy in the ChatGPT-5 study, so add or correct items.
• Edit the weight. In Tabaq you can change the weight or ingredients after the analysis and the calories and macros update.
• Add what the camera cannot see: cooking oil, sauce, sugar in a drink.
• Weigh a few staples once (rice, pasta, olive oil, bread) so you know what your usual portion looks like.
• Keep logging consistently. Random errors partly cancel across meals, while a consistent bias is easier to spot and correct when you track weight over a few weeks.

When to weigh instead

Use a scale if you are tightly managing calories or macros, cooking for a medical diet, or if a meal has no visible structure (smoothies, soups). If you use insulin, have diabetes or a history of disordered eating, talk to your doctor or dietitian about how to track.

Try it on your own plate

Tabaq analyses your meal photo, and during the trial you get 3 days of unlimited photo analysis. Photograph your next meal and compare the estimate with what you expected, then edit the weight if it looks off.