How accurate are AI calorie estimates?

How photo estimates work

A vision model identifies the foods on the plate, guesses the portion size from visual cues like plate size and depth, then maps each item to nutrition data. Each step can add error, and portion size is usually the biggest one.

Where AI is accurate

These cases look a lot like the data the models were trained on, so the guesses are usually close.

• Single, whole foods: a banana, an egg, a chicken breast.
• Plated meals where each component is visible and separate.
• Packaged foods when you include the label or brand in the message.

Where it gets things wrong

• Hidden fat: oil, butter and cream used in cooking are invisible in a photo.
• Mixed dishes: curries, casseroles, smoothies and burritos hide what is inside.
• Portion depth: a bowl of rice photographed from above can be shallow or deep.
• Dressings and sauces poured on top.

Compare it to the alternatives

Manual logging isn't exact either. In one classic study, people who said they couldn't lose weight underreported their intake by almost half. US rules also allow nutrition labels to be off by up to 20%, and restaurant meals often contain more than stated.

So the real question is not whether AI is perfect but whether it is consistent and easy enough that you keep logging. A tool you use every day beats a precise one you quit after a week.

How to get better estimates

• Shoot from a 45 degree angle so depth is visible, with the whole plate in frame.
• Add a caption: cooking method, oils, sauces, or "I ate half".
• Correct it once. Saying "that was 200 g of rice, not 150" teaches you what portions look like.
• Watch your weekly weight trend. If it doesn't match the logged deficit, add a buffer of 10 to 15%.