On this page
- Are AI calorie counters accurate?
- What are the three separate errors in a photo estimate?
- Does correct food recognition mean the estimate is right?
- Are AI photo calorie counters accurate for mixed dishes and hidden oil?
- How much does each error change protein and calories?
- Does AI calorie tracker accuracy improve with better photos or extra detail?
- What should you check before saving an AI estimate?
- How can you test one ordinary meal against measured ingredients?
- Who should skip calorie counting from photos?
- What should you do next?

Are AI calorie counters accurate?
Are AI calorie counters accurate? They give estimates, not measurements. One 2026 test found the photo features of 4 apps ran low on calories (kilocalories, shortened to kcal below). That research abstract, from the NIH Intramural Research Program, was published in a journal supplement in July 2026 by Charles and colleagues. Their meals were planned from national food composition tables for a controlled feeding study. Photographs of 102 prepared meals were run through the AI features of 4 commercially available apps, and each result was compared with the planned values.
The meals averaged 918 kcal, 36 g of protein and 61 g of fat. All 4 apps underestimated energy, with mean differences between −252 and −345 kcal per meal, about 33% in the authors’ summary. All 4 underestimated fat by about 30 g. Two of the apps underestimated carbohydrates by about 14 g, and the other two did not differ significantly.
Averages hide the spread. For one app, the limits of agreement, the range expected to hold about 95% of individual differences, ran from 884 kcal too low to 246 kcal too high. The authors also report that consistency, measured with an intraclass correlation, was moderate to good for carbohydrates and poor for the other outcomes. The abstract lists protein only as the meals’ average content, not as an error, so this test does not tell you how far off protein was.
The study has limits. It is an abstract with brief methods, it does not give the app versions, and its meals were planned for one controlled feeding study. The authors note that the size of the underestimate is similar to what has been reported for self-report diet assessment methods, so underestimating is not unique to AI. One test does not show that every photo feature runs low, and the 2024 study below found errors in both directions.
What are the three separate errors in a photo estimate?
A photo estimate combines three steps, and each can fail on its own: recognition, portion and nutrient calculation. Treating them separately shows what you can fix when you review a result.
| Step | What goes wrong | What you can check |
|---|---|---|
| RecognitionWhat is on the plate? | A wrong food, or a missing component such as a sauce, spread or dressing. | Read every item name. Add what the photo cannot show. |
| PortionHow much is there? | A photo shows area, not depth or weight. A heap and a thin spread can look alike. | Compare the amount with a package weight, a measuring cup or a scale. |
| Nutrient calculationHow much protein and energy does that amount carry? | The wrong database entry: raw instead of cooked, lean instead of regular, or a similar product with different numbers. | Compare with the package label or USDA FoodData Central, in the same raw or cooked state. |
In practice, an app can name every food correctly and still land far off, because a portion or database error does not depend on recognition. Calorie error and protein error also differ, as the worked example below shows.
Does correct food recognition mean the estimate is right?
Correct recognition does not settle accuracy. In 2024, Li and colleagues in Nutrients photographed 22 foods and drinks, taken from Western and Asian sample diets, in controlled lighting and uploaded them to 7 apps with AI image recognition, using the app versions listed in the paper. The images held 39 food components in total. The best app identified 38 of 39 (97%), while the lowest-scoring apps identified 18 of 39 (46%).
Identification was not the end of the problem. Four of the 7 apps gave energy values automatically. For single foods, their mean differences from the reference food records ranged from 47% too low to 44% too high. For mixed dishes the authors reported notable discrepancies: one app underestimated eggs on toast with butter by 35% because it left out the butter, and another underestimated the same image by 73%. For two Asian dishes, a noodle soup and a milk tea, apps failed to identify the components, with energy up to 76% too low.
The study has limits. It tested free app features, took photos in controlled lighting at a fixed angle, and used only 22 images, so everyday photos may do better or worse. The authors say mixed dishes and culturally varied foods need better training.
Are AI photo calorie counters accurate for mixed dishes and hidden oil?
The evidence points to the weakest results where a photo hides ingredients: mixed dishes, cooking fat and sauces. A 2026 benchmark in Scientific Reports by Vedovelli and colleagues tested 40 general-purpose vision-language models (AI models that read images and text), using the October 2025 versions, on 50 laboratory-photographed dishes with reference nutrient values, chosen across 5 calorie bands. In the 10 hardest dishes, fat errors were the largest, protein errors intermediate and carbohydrate errors the smallest. The authors link the pattern to hidden oils, sauces and fatty meat that a 2D image cannot show, and describe depth and volume estimation from a single flat image as inherently ill-posed. They also found that energy errors tracked fat errors closely, and read this as hidden fat, more than portion misjudgment, driving the calorie errors in their dishes.
Four sources of error recur, and you can look for each one in your own photos:
- Mixed dishes. Stews, bowls, casseroles and wraps hide how much of each ingredient is inside. The 2024 study above found apps often failed to split a mixed dish into its parts.
- Hidden oil and butter. Cooking fat soaks into food or sits under it, and the 2024 study’s butter omission shows what happens when an app skips it. Our guide to hidden calories lists the usual suspects.
- Depth. A photo shows the surface. A deep bowl and a shallow plate can cover the same area while holding very different amounts, which is why the same photo can fit several portion sizes.
- Serving ambiguity. “A bowl of rice” has no fixed weight, and a database entry may describe cooked food, raw food or a packaged serving. USDA FoodData Central (SR Legacy, checked November 2026) lists 85 g of raw skinless chicken breast at about 19 g of protein and 85 g of roasted chicken breast at about 26 g.
Error also exists before a camera is involved: how accurate calorie labels and apps are covers label rules and database differences.
How much does each error change protein and calories?
Each error moves protein and calories differently, so the check you need depends on what you track. The example below is a plain meal calculated from USDA FoodData Central values: roasted chicken breast, cooked white rice, boiled broccoli and olive oil. It is a calculation, not a kitchen test, and it is not output from any app.
| Ingredient, cooked weight | Protein | Energy |
|---|---|---|
| Chicken breast, roasted3 oz (85 g) | 26 g | 140 kcal |
| White rice, long-grain, cooked1 cup (158 g) | 4.3 g | 205 kcal |
| Broccoli, boiled, no salt½ cup (78 g) | 1.9 g | 27 kcal |
| Olive oil1 tbsp (13.5 g) | 0 g | 119 kcal |
| Whole meal | 32 g | 492 kcal |
The next table changes one thing at a time and recalculates the whole meal. Each row stands for an error a photo estimate could make; the rows are scenarios, not measured app results.
| One change to the meal above | Protein | Energy |
|---|---|---|
| Oil left out, because it is not visible | 32 g | 373 kcal |
| Twice the oil, 2 tbsp (27 g) | 32 g | 612 kcal |
| Chicken portion about 50% larger, 4.5 oz (128 g) | 46 g | 563 kcal |
| Raw chicken entry used for the cooked 3 oz (85 g) | 25 g | 454 kcal |
| Rice portion 1½ cups (237 g) | 35 g | 595 kcal |
Leaving the oil out removes about 24% of the meal’s calories and none of its protein. A larger chicken portion or the wrong raw or cooked entry shifts protein by several grams. This suggests that, if you track protein first, the checks that matter most are the identity and weight of the protein food. If you also track calories, the oil and the starch portion deserve a second look. For no-scale ways to size a protein portion, see estimating protein without a food scale.
Does AI calorie tracker accuracy improve with better photos or extra detail?
In one 2026 benchmark, the choice of model mattered far more than extra photos or ingredient notes. That benchmark tested general-purpose models and not finished tracker products. Among the settings the authors varied (model, dish sample and temperature), model architecture explained 99.6% of the variance in performance. Adding more views, up to 5 images per dish, gave no significant benefit. Giving the models a text list of the dish’s ingredients, without amounts, improved the combined error score by about 5.5% on average, with results ranging from 19% better to 3% worse depending on the model. Smartphone photos taken by a nutritionist scored better than the laboratory images, although the two sets held different dishes, so the comparison is not like for like.
The benchmark also compared the models with one nutritionist who estimated protein, fat, carbohydrate and energy from photos of 7 dishes. Protein was the nutritionist’s most accurate nutrient, with a mean absolute error of 1.6 g, and the AI protein errors in that comparison were more than 5 times larger in absolute terms. The authors call the comparison indicative, because the nutritionist and the models saw different dishes. They conclude that current models suit consumer calorie tracking, while clinical-grade macronutrient estimates will need better models, and they note that no model reached the ±10% accuracy that the Academy of Nutrition and Dietetics recommends for dietary recalls.
These results describe model versions from October 2025 and benchmark dishes, not any one app as of November 2026. Technology changes quickly, so the next two sections show how to check your own app.
What should you check before saving an AI estimate?
A short review before you save catches most of the errors in the table above. Work through these in order:
- Name. Read the food name. Change it if the app picked a similar but different food, such as tofu for chicken or a sweetened yogurt for plain.
- Missing items. Add what the camera cannot see: oil, butter, dressing, sauce, cheese, a drink.
- Raw or cooked. Match the entry to the state of the food on your plate.
- Portion. Compare the amount with something you know, such as a package weight, a measuring cup or a no-scale portion guide. Edit the grams, not just the protein number.
- Protein number. Check that it is plausible for the food and portion. Our high-protein foods chart gives USDA values by serving for comparison.
- Packaged foods. Use the package. A barcode or label match gives you the manufacturer’s declared values, which is usually a better starting point than a photo guess, and reading a nutrition label for protein shows how to match the serving.
- Reuse. Save the corrected meal if you eat it often, so next time starts from a checked entry.
How can you test one ordinary meal against measured ingredients?
You can measure your app’s error on one meal you eat often, and the result is more useful than any average from a study. These steps take about one meal and a kitchen scale:
- Pick a simple meal. Choose one with 3–5 ingredients that you can weigh separately.
- Weigh each ingredient. Note whether the weight is raw or cooked, and measure oil and butter by the teaspoon or tablespoon.
- Photograph the plate as you normally would. Write down the app’s food names, portions, protein and calories before you edit anything.
- Calculate the measured total. Use package labels or USDA FoodData Central, with entries that match your raw or cooked weights.
- Compare layer by layer. Did the app name every component? Were the portions close? Did the entries match the weights you used? Then compare protein and calories.
- Repeat with 2 more meals. Include one mixed dish and one meal cooked with oil, because those are where photos struggle.
Decide your own tolerance in advance, for example how many grams of protein off you are willing to accept for a meal. If the app misses in the same direction each time, you can adjust the habit, such as always adding the oil. This is a personal calibration based on a few meals, not a validation of the app. For a non-photo method, see the guide to tracking protein without counting calories. How to keep a food diary shows what to record and how to review your meals.
Who should skip calorie counting from photos?
Calorie counting is optional, and photo estimates are not required for protein tracking. If you have a history of an eating disorder or disordered eating, skip calorie counting and talk to a clinician. Is counting calories healthy? discusses who benefits and who does not, and the protein tracking guide offers a gentler option. If you take GLP-1 (glucagon-like peptide-1) or diabetes medicines, ask your prescriber before cutting calories. If you are pregnant or breastfeeding, are under 18, have kidney disease or follow a medical diet, talk to your care team before you change what you eat or how you track it. If a clinician has set a calorie, carbohydrate or protein target for you, ask whether an AI photo estimate is accurate enough for your plan.
What should you do next?
Pick one meal you eat often, weigh it once and compare it with your app using the steps above. For other apps and features to compare, see how to choose a protein tracker for iPhone; if you do count calories, how to count calories covers the routine, and the calorie calculator gives a starting estimate. For protein, start from the daily protein calculator.
YolkPal is an iPhone protein tracker, coming soon to the App Store, with calories optional. You can take a photo of the plate for a protein estimate and edit the food, the portion and the protein, and the calories if you track them, before saving. AI results are estimates, so check them with the steps above.
Sources
- Charles O, et al. Photograph-Based AI Features in Calorie-Tracking Apps Underestimate Energy Content of Meals. Curr Dev Nutr 2026;10 (Supplement 1):108738. Research abstract in a journal supplement. doi:10.1016/j.cdnut.2026.108738
- Li X, et al. Evaluating the Quality and Comparative Validity of Manual Food Logging and Artificial Intelligence-Enabled Food Image Recognition in Apps for Nutrition Care. Nutrients 2024;16(15):2573. doi:10.3390/nu16152573
- Vedovelli L, et al. Model architecture dominates nutritional estimation accuracy in vision-language systems. Sci Rep 2026;16:29227. doi:10.1038/s41598-026-58755-w
- USDA FoodData Central, SR Legacy: chicken, broilers or fryers, breast, meat only, cooked, roasted (FDC 171477); chicken, broiler or fryers, breast, skinless, boneless, meat only, raw (171077); rice, white, long-grain, regular, enriched, cooked (168878); broccoli, cooked, boiled, drained, without salt (169967); oil, olive, salad or cooking (171413). fdc.nal.usda.gov
This article is general information, not medical advice. If you have a health condition, are pregnant, or follow a diet set by your doctor, talk to them or a registered dietitian before changing how you eat.


