How accurate is AI calorie counting?

Where the error comes from

• Portion: a photo does not reliably convey depth or bowl size. 150 g and 250 g of rice look nearly identical from the same angle.
• Invisible fat: sautéing oil and salad dressing carry much of the energy and are not visible.
• Mixed dishes: in a stew or a soup the ingredients sit inside one another, so the composition can only be guessed.
• Regional recipes: the same dish name can mean noticeably different ingredients from one kitchen to the next.

How to get more out of the estimate

• Correct the amount: if the app allows it, don’t accept the estimate — type the quantity you know.
• Use the same bowl: eat from the same plate daily and the error becomes consistent, which keeps the trend honest.
• Say the fat out loud: a note like “cooked with two tablespoons of olive oil” supplies what the image cannot.
• For packaged food, use the label: there is no reason to estimate when a number is printed on the box.
• Judge by the weekly average, not by one meal. That is the number you can actually act on.

The one feature to insist on

Correction. An app that shows you a result and offers nothing but “OK” loses your trust at the first bad estimate, and then it gets deleted.

The right behaviour is: an estimate arrives, you change the amount, and calories and macros are recalculated in the same proportion. The rescaling has to be done from the ORIGINAL values each time — otherwise someone who goes 200 → 100 → 200 ends up with numbers that no longer add up.

What calories alone don’t tell you

Two days with the same calorie total can look very different: protein distribution, meal timing, sleep and movement are not independent of one another. Calories are an input, not a report card.

This is also why calorie tracking on its own tends to last a few weeks and then stop. Making it part of a system that sees the rest of the day holds up longer.

This page is for general information and is not medical advice. Talk to your doctor before making decisions about nutrition, training or supplements.

How it works in Testoman

You add a meal by photo or in writing, and the analysis comes back item by item. The gram field is optional: leave it empty and the estimate is used, fill it in and the model works from exactly that amount.

You can also correct a meal you saved earlier — change the amount and the macros rescale, with the day’s total updating immediately. No second analysis is spent.

If you have declared dietary restrictions (lactose, vegan, halal and so on), a warning band appears on the analysis result. It is not a safety guarantee and does not replace reading the label.