How Accurate Are AI Calorie Counter Apps? What Tests Show

This article is for educational purposes only and is not medical advice. Talk with a qualified healthcare professional before making major nutrition, weight loss, or medication-related changes.

AI calorie counters that work from a meal photo tend to log too few calories when you do not edit the estimate. In a 2026 test by National Institutes of Health researchers, four popular apps logged 102 precisely weighed meals about a third too low, by 250 to 345 calories per meal on average. That result is a conference abstract and has not been peer reviewed.

Earlier studies of image-based food recognition found average calorie errors anywhere from 0.10% to 38.3%, depending on the system and the test images. No single figure describes every app or every meal. This guide covers what was tested, what the numbers do and do not show, and how to correct a photo estimate.

Disclosure: Kalo publishes this site and makes a calorie tracking app. Kalo was not part of any study described here, and none of these numbers describe Kalo.

Key Takeaways

Unedited photo estimates ran low. Four apps logged weighed meals about 250 to 345 calories low on average in the 2026 test.
Fat was a large part of the miss. All four apps logged about 30 grams less fat per meal than the meals contained.
No single figure fits every app. A review of 52 studies found average calorie errors from 0.10% to 38.3% across different systems.
The app test is preliminary. It is a conference abstract, and it measured photo estimates with no manual correction.
Treat the estimate as a draft. Check the portions and add oils, dressings, and sauces before you save.

How Accurate Are AI Calorie Counter Apps in a Direct Test?

In the 2026 test, every app logged fewer calories, on average, than the meals contained. Researchers at the National Institutes of Health took standardized photos of 102 meals from a research kitchen where each ingredient is weighed to the nearest 0.1 gram. They ran the photos through MyFitnessPal, Lose It!, Cal AI, and Appediet.

According to the American Society for Nutrition's summary of the abstract, all four apps underestimated calories by about 250 to 345 per meal on average, or about a third, and fat by about 30 grams. Healio's report from the conference lists the average shortfall for each app.

Average calories logged below the weighed value per meal, across 102 research meals. Preliminary conference data, as reported by Healio.
App	Calories low per meal	95% confidence interval
Appediet	252	210 to 295
MyFitnessPal	327	269 to 385
Lose It!	333	282 to 383
Cal AI	345	296 to 392

This is not a ranking. The differences between the apps are smaller than the gap between every app and the weighed meals. Read the table with these limits in mind:

It is preliminary. Abstracts at this conference are selected by a committee but have not generally gone through the peer review required for journal publication.
It measured the photo estimate alone. The lead researcher said people who log a photo without adjusting portions or entering amounts should "take the results with a grain of salt."
The meals came from a diet study. They were not home or restaurant plates. Early follow-up results suggest the apps struggled more with low-carb ketogenic meals, which are higher in fat.

Two details help explain the shortfall. All four apps estimated carbohydrates more consistently than the other macronutrients. And at 9 calories per gram, 30 grams of fat is about 270 calories, which is in the same range as the calorie shortfall itself.

What Does Earlier Research on Photo Calorie Estimates Show?

Earlier research does not support one accuracy number. A 2023 systematic review of image-based dietary assessment tools covered 52 studies published from 2010 to 2023. Average calorie errors ranged from 0.10% to 38.3%, and average volume errors ranged from 0.09% to 33%. The studies used different image sets and reported results in different ways, so the authors could not combine them into one figure.

Two points from that review matter when you read a claim about an app:

It reviewed published research systems. Apps on the market were not evaluated unless a validity test had been published.
Simple images did better. Images of one simple food, such as an apple, had lower errors. Mixed dishes, such as curries, carried a higher risk of identification errors.

The review concluded that the tools needed more development before use as stand-alone dietary assessment methods in nutrition research or clinical practice. The table below separates the three kinds of evidence this guide uses.

What each source tested and its main limit. The three results are not directly comparable.
Evidence	What was tested	Calorie result	Main limit
2026 app test, conference abstract	Photo features of four consumer apps on 102 weighed research meals	About 250 to 345 calories low per meal, about a third	Not peer reviewed. Photo estimates only, with no manual edits.
2023 systematic review of 52 studies	Research systems on test image sets	Average errors from 0.10% to 38.3%	Not consumer apps. Results could not be combined.
2021 Nutrition5k paper	A research camera rig on 5,066 cafeteria dishes	Average error of 26.1% from one image, 16.5% with depth data	Not a phone app. Dishes averaged 255 calories.

How Accurate Is SnapCalorie?

The sources reviewed for this guide contain no independent accuracy figure for SnapCalorie, and SnapCalorie was not one of the four apps in the 2026 test. SnapCalorie's website says its computer vision is twice as accurate as nutritionists and links to Nutrition5k, a 2021 computer vision paper.

That paper tested a research setup, not the phone app. Dishes from campus cafeterias were captured on a custom rig with a depth camera and a scale under the plate. The model's average calorie error was 26.1% from a single image and 16.5% when depth data was added. The comparison with people involved 4 nutritionists and 16 other raters judging 10 plates from photos, and the authors describe it as not an exhaustive study.

Cal AI was in the 2026 test. For the wider record on that app, see our review of the evidence on Cal AI. For a general chatbot rather than a dedicated app, see whether ChatGPT can count calories.

Why Do Photo Estimates Miss Calories?

A photo shows the surface of a meal, and much of the energy is not on the surface. The evidence points to three reasons.

Fat is hard to see. Oil, butter, dressing, and cheese add calories without changing how a plate looks. The tested apps logged about 30 grams less fat per meal than the meals contained, and early results were worse for higher-fat ketogenic meals.
Portion size is inferred from a flat image. In the Nutrition5k paper, adding depth data lowered the average calorie error from 26.1% to 16.5%. That suggests a large share of the error comes from judging volume.
Mixed dishes hide ingredients. The 2023 review found lower errors for single simple foods and a higher risk of errors for mixed dishes.

Are AI Calorie Counters Accurate Enough for Weight Loss?

On the current evidence, not without correction. A food log that runs low can hide all of a planned calorie deficit. In the 2026 test the average miss ran in one direction, so it does not even out from meal to meal.

The table shows the arithmetic. It is an illustration, not a prediction of weight change. If a log reads 2,000 calories and runs low, the actual intake is the logged number divided by one minus the shortfall.

Illustrative arithmetic: actual intake = logged calories ÷ (1 − shortfall). Applying one average shortfall to a whole day is an assumption.
Shortfall in the log	Log reads	Actual intake	Calories not in the log
10%	2,000	About 2,220	About 220
20%	2,000	2,500	500
33%	2,000	About 2,985	About 985

A 20% shortfall on a 2,000-calorie log is 500 calories, which would cancel a planned 500-calorie daily deficit. Two cautions keep this in proportion. First, the test covered photo estimates only, not barcode scans or typed entries, so a whole day of mixed logging may differ. Second, logging by hand runs low too. In a 1992 study of self-reported and actual calorie intake, 10 adults with obesity who reported being unable to lose weight despite dieting under-reported what they ate by an average of 47%.

The practical check is your own trend. If your weight over several weeks does not match what your log implies, check whether the log is running low before you cut calories further. Our guide to not losing weight in a calorie deficit walks through the checks in order.

How to Get a More Accurate Estimate From an AI Calorie Counter

Treat the photo result as a first draft, then correct it. These steps follow from where the tested apps went wrong.

Check each portion before you save. Compare the logged amount with what is on the plate and change it. The researchers' caution applied to photos logged without adjusting portions or entering amounts.
Add the fats the camera cannot see. Enter cooking oil, butter, dressings, sauces, and cheese as separate items. Missing fat was a large part of the shortfall in the 2026 test. See how to track cooking oil calories.
Use the label for packaged food. Scan the barcode or copy the Nutrition Facts panel, then scale it to the amount you ate. Our food label guide covers serving sizes.
Enter recipes by ingredient. For homemade dishes, total the ingredients you used and log your share. See how to calculate calories in homemade food, or use the recipe nutrition calculator.
Weigh calorie-dense foods when the number matters. A kitchen scale removes the portion guess for foods such as oils, nut butters, rice, and pasta.
Compare the log with your weight trend. Review several weeks, not single days.

Where Kalo Fits

Kalo is a calorie and macro tracking app for iOS and Android. Its App Store listing describes four ways to log a meal: a photo, a voice or text description, or a barcode scan. The listing says photo and voice entries use AI to create an editable estimate, and that you can adjust foods, servings, calories, and macros before saving.

Kalo was not one of the four apps in the 2026 test, and Kalo has not published a validation study. Nothing on this page measures Kalo. The same rule applies to a Kalo estimate as to any other: review it, correct the portions, and add what a photo or a short description can leave out.

Who Should Be Careful With Calorie Tracking Apps?

Calorie tracking is a tool, not medical care. If you have a history of disordered eating, are managing a medical condition, or find that tracking food makes you anxious, talk with a qualified clinician or registered dietitian before relying on an app's numbers.

Frequently Asked Questions

Which AI calorie counter is most accurate?

No published evidence identifies one. In the 2026 test all four apps logged meals low, and the differences between them were smaller than the gap between every app and the weighed meals. If you are choosing an app, our comparison of calorie counter apps covers logging methods and what each listing says is free or paid.

Is a photo or a typed description more accurate?

The sources reviewed here do not compare them. The 2026 test covered photo estimates only. A description lets you state amounts and hidden ingredients, such as two tablespoons of olive oil, but the result is still an estimate to review.

Is any calorie count exact?

No. U.S. Nutrition Facts labels round calories to the nearest 5 up to 50 calories and to the nearest 10 above that, under 21 CFR 101.9. The same regulation treats a label as out of compliance when lab-measured calories are more than 20% above the declared value. Every food log is an estimate. Aim for one that is consistent and that you have checked.

Sources

Sources checked October 6, 2026. The 2026 app test is a conference abstract that has not been peer reviewed, and its per-app figures come from news coverage of the presentation. The calorie table in this guide is illustrative arithmetic. Kalo publishes this guide and makes a calorie tracking app. No clinical review is claimed.