免費試用 Nutrify 7 天 - 全部功能開放免費開始
Nutrify 標誌Nutrify

How Accurate Is AI Calorie Counting?

By Nutrify Team•Published on September 1, 2026•Updated October 29, 2026•AI Food Tracking#ai-tracking#accuracy
How Accurate Is AI Calorie Counting?

AI calorie counting from a photo is a useful estimate, not a measurement. In independent tests, photo-based estimates were off by roughly a quarter to a third, and they usually came in too low. An NIH test of four popular apps found them 250 to 345 calories short per meal.

Accuracy improves a lot when you tell the app what's in the dish and how much, and when you use a barcode or a scale for the foods you eat most.

Key takeaways

  • Errors lean low: most studies found AI underestimates calories, especially for big portions and high-fat meals.
  • Identifying food is easier than weighing it: the AI often names the dish correctly but misjudges the amount and the hidden oil.
  • Context helps: adding ingredients and amounts made estimates more accurate in a 2025 ChatGPT study.
  • People aren't perfect either: memory-based food diaries have missed up to half of real intake.
  • Check against your scale: a 2-4 week comparison of logged calories and weight change shows whether your logs run low.

What the studies found

StudyWhat was testedMain result
NIH (NIDDK), presented July 2026MyFitnessPal, Lose It!, Cal AI and Appediet on photos of 102 meals with known nutritionCalories 250-345 too low per meal on average; fat about 30 g too low; high-fat keto meals hardest. Preliminary, not yet peer-reviewed
Fridolfsson and colleagues, 2025ChatGPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro on 52 standardized food photosMean calorie error 35.8% for ChatGPT and Claude; Gemini 64-110% across nutrients; all underestimated larger portions
Rodríguez-Jiménez and colleagues, 2025ChatGPT-5 on 195 dishes, with more and more contextError dropped as descriptions and ingredient amounts were added; removing the photo made it worse
Nakagawa and Yamamoto, 2026Three Claude models on 691 Japanese and 1,463 US meal photos and about 1,000 Japanese packaged-food imagesClaude Sonnet 4.6 was off by 117 calories per US photo on average; packaged foods were easier than meals; salt was poorly estimated
Vedovelli and colleagues, 202640 vision-language models vs professional nutritionistsThe model used mattered most; extra photo angles didn't help; nutritionists were far more accurate, especially for protein
Serra and colleagues, 2026One AI photo app (SNAQ) in daily life for 7 days, vs doubly labelled water, 20 women with obesityApp underestimated intake by 25% (817 calories a day); a 24-hour recall interview underestimated by 50%

Two patterns run through all of them. Errors are big enough to matter for a single meal, and they mostly point in one direction: too low.

Why AI calorie counts come out low

A camera sees the top of your plate. It can't weigh the food, see how deep the bowl is, or tell how much oil went into the pan. Calories hide in exactly those places.

  • Portion size: the same photo can show 150 g or 250 g of pasta. In the 2025 study, underestimation grew as portions got bigger.
  • Hidden fat: butter, cooking oil, dressings and creamy sauces are calorie-dense and hard to see. In the NIH test, fat was the most underestimated nutrient and high-fat meals were the hardest.
  • Mixed dishes: stews, curries, casseroles and burritos hide most of their ingredients inside.
  • Unusual or regional foods: a model that has seen fewer examples of a dish has less to go on.

Packaged foods are the exception, because the label does the work. That's why scanning a barcode beats a photo for anything that comes in a wrapper. For more on how apps estimate amounts, see how AI handles portion size estimation.

AI vs manual logging vs memory

AI isn't competing with perfection. Every way of counting calories has error:

  • Memory: in a classic study of people who said they couldn't lose weight on under 1,200 calories, they under-reported their real intake by 47% and over-reported exercise by 51% (Lichtman and colleagues, 1992). In the 2026 SNAQ study, a structured 24-hour recall missed 50%, worse than the app's 25%.
  • Food labels: the FDA treats a packaged food as misbranded only when its calories are more than 20% above the label, so small overages are allowed.
  • Restaurant numbers: in lab tests of 269 restaurant foods, stated calories were right on average, but 19% of items had at least 100 calories more than listed (Urban and colleagues, 2011).
  • Weighing and logging by hand: still the most accurate home method, but it takes the most time.

That makes a photo estimate less accurate than a weighed, barcode-based log, and in some studies more accurate than memory. For a side-by-side of the methods, see AI vs manual calorie logging and barcode scanning vs AI photo recognition.

Is AI calorie counting accurate enough for weight loss?

It can be, as long as you correct for the direction of the error. Random error, where some meals come in high and some low, mostly cancels out over a week. A steady undercount doesn't.

Say your target leaves a 500-calorie daily deficit. If your logs undercount three meals by 250 calories each, you're really eating 750 calories more than you think, and the deficit is gone. That's the kind of gap the NIH test found.

The fix is a simple check:

  1. Log everything for 2-4 weeks and weigh yourself several times a week at the same time of day.
  2. Compare your average logged intake with what your weight trend says. If you're logging a 500-calorie deficit but your weight isn't moving, your logs are probably running low.
  3. Adjust: either log more carefully (see below) or lower your calorie target by the gap you found.

Your TDEE is the starting point for that comparison.

How to make AI calorie tracking more accurate

  • Shoot from above, with the whole plate in frame. A fork or your hand in the shot gives a sense of scale.
  • Add what the camera can't see. Type or say "cooked in 1 tablespoon of olive oil" or "with 2 tablespoons of ranch". Context improved estimates in the ChatGPT-5 study.
  • Give amounts when you know them. "200 g chicken breast" beats a guess every time.
  • Correct the portion before you save. If the estimate looks small for what you ate, it probably is.
  • Use a barcode for packaged food and a kitchen scale for the foods you eat daily, like oats, rice, nut butter and cooking oil.
  • Separate mixed plates when you can, or log the main ingredients of a stew rather than "stew".

Our guide to tracking calories accurately has more on restaurant meals and eating out.

When to use something other than a photo

  • Packaged foods: scan the barcode.
  • Foods you eat every day: weigh them once, save them, and reuse the entry.
  • Medical diets: if you manage diabetes, kidney disease or another condition through food, a photo estimate isn't precise enough. Work with your doctor or a registered dietitian.

Where Nutrify fits

Nutrify gives you photo, barcode and voice logging in one place, so you can use a photo for a quick plate and a barcode or a typed amount when you want precision. Every estimate can be edited before you save it. Nutrify has a 7-day free trial, then a paid subscription. If you're comparing apps, see Cal AI vs MyFitnessPal.

Download on the App StoreGet it on Google Play

Sources

常見問題

尋找常見問題的解答

  • In independent tests, photo-based estimates were off by roughly a quarter to a third, and they usually came in too low. An NIH test of four popular apps found them 250-345 calories too low per meal, and a 2025 study found ChatGPT and Claude off by about 36% on standardized photos.

  • Usually, yes. Studies found errors that lean low, and the gap grows with bigger portions and fattier meals. In one real-life test against doubly labelled water, a photo app missed about 25% of what people actually ate.

  • It can work if you check and correct the estimates. A steady undercount of a few hundred calories a meal can erase a 500-calorie deficit, so compare your logs with your weight trend over two to four weeks and adjust.

  • Roughly. In a 2025 study of 52 standardized food photos, ChatGPT-4o and Claude 3.5 Sonnet both had a mean error of 35.8% for calories, and both underestimated larger portions. Telling the model the ingredients and amounts improves the estimate.

  • Not automatically. Photo logging is faster, but weighing food and logging it by hand, or scanning a barcode, is more accurate. Memory-based logging is the weakest: people in one classic study under-reported by 47%.

  • A photo shows the food's surface, not its weight, depth or the oil and butter inside it. That's why portion size and hidden fat cause most of the error, even when the food is identified correctly.

  • Take the photo from above with the whole plate in frame, add the ingredients and amounts you know, log cooking oil and sauces separately, and use a barcode or kitchen scale for foods you eat often.

  • No consumer photo app is validated for medical nutrition care. A 2026 study of one AI app against doubly labelled water found poor agreement for individuals. If you manage a condition through diet, work with your doctor or a dietitian.

Nutrify Team's avatar

Nutrify Team

A team of nutrition experts, developers, and health enthusiasts dedicated to making healthy living accessible to everyone.

Track this in Nutrify

Free to download on iOS and Android. Everything below is part of the app.

在 Google 上追蹤 Nutrify

將 Nutrify 加入偏好來源,在 Google 搜尋結果中看到更多我們的營養指南。