Myth 1 of 3

Can AI count calories from a photo?

In 2025, researchers gave ChatGPT-5 photos of 195 dishes: recipe photos, a public meal-photo dataset, and home meals weighed in a kitchen. They compared its estimates with the known values, then gave it more context, one step at a time.

For protein the gap was bigger: 42.6% average miss from the photo alone, 20.9% with detailed ingredients. "Average miss" here is the average size of the error, whether the estimate was too high or too low.

Why the photo misses

The camera sees the top of the plate. It does not see:

• How much fat went in. One tablespoon of olive oil is 119 kcal and can look like a shine.
• Depth. A deep bowl and a shallow plate photograph the same.
• What is under the sauce. Or how much sauce there is.
• Raw or cooked. Same food, very different numbers per 100 g. See myth 3.

On a 600 kcal meal, a 30% miss is 180 kcal, up or down, and the photo will not tell you which. In the checked example in our kit, the first photo estimate was 154 kcal above the corrected estimate.

What to do instead

1. Give the amounts you know. "Chicken 120 g cooked, rice about 150 g, 2 teaspoons of oil." The study tested detailed ingredient lists; giving rough amounts is our practical advice, not a tested result.
2. Make it list its guesses before any totals. Then you can fix the biggest one.
3. Fix the biggest item first. In our kit example the rice portion moved the total most (65 kcal), then the chicken (50 kcal) and the oil (40 kcal).

A starter line to paste after your photo:

Before giving any totals, list every item you see with the weight in grams and the oil or butter you assume. Mark each one "confident" or "guess". Then stop and wait for my corrections.

The limits of this answer

One study, one model, 195 dishes. Other assistants and newer versions may do better or worse. And an average miss means some meals were missed by much more.