LogmaCal AI is roughly accurate for simple, visible foods, but a 2026 study from the National Institutes of Health found that popular photo calorie apps, Cal AI among them, undercounted real meals by about 250 to 345 calories each, almost all of it fat the camera cannot see. So the honest answer is this: close enough for a plain chicken breast or a piece of fruit, often hundreds of calories low for anything cooked in oil, sauced, or mixed together.
That gap matters because 300 calories a day, counted the wrong way, is roughly the difference between losing weight and standing still. Below is what the testing showed, when you can trust a photo, and the one thing you can do to make any calorie app land closer.
What the NIH study actually found
In July 2026, researchers at the National Institute of Diabetes and Digestive and Kidney Diseases, part of the NIH, tested four popular calorie apps: MyFitnessPal, Lose It!, Cal AI and Appediet. They photographed more than 100 meals made in a controlled metabolic kitchen, where every ingredient was weighed, so the apps' guesses could be checked against the real numbers.
Across all four apps, the photos came back about 250 to 345 calories too low per meal, roughly a third of the real total. Fat was the big miss: the apps underestimated it by around 30 grams a meal. Carbohydrates were the one thing they got reasonably consistent. A follow up look at more than 200 extra meals found the apps did even worse on low carb, higher fat plates, which fits the pattern.
One caveat worth saying out loud: this was presented at a conference (NUTRITION 2026) and has not been through full peer review yet, so treat it as strong early evidence, not the last word. But it lines up with what anyone who has logged a greasy takeout dinner already suspected.
Why photos miss calories: the fat you can't see
A camera reads shape, color and size. It does not read what soaked in. Fat carries about 9 calories a gram, more than twice what carbs or protein carry, so the invisible stuff is exactly the stuff that moves your total the most.
Thirty grams of fat, the amount the NIH apps missed on average, is about 270 calories. That is one extra tablespoon of oil in the pan, the butter the restaurant finished the steak with, the dressing already tossed through a salad, the oil a falafel or a piece of döner was fried in. None of it changes the photo much. All of it changes the calories a lot.
This is not really a knock on Cal AI's model. It is a limit of the input. A photo of a grilled chicken Caesar and a photo of the same salad with three tablespoons of ranch look almost identical, and they are about 300 calories apart.
When Cal AI is accurate enough
Photos are not useless. They are accurate enough when what you are eating is simple and the calories are mostly on the surface. Here is a rough guide to where a photo lands.
| What you're logging | How a photo tends to do |
|---|---|
| An apple, a banana, plain Greek yogurt | Good. Little hidden fat, familiar portions |
| A packaged bar or a labeled snack | Skip the photo, scan the barcode or read the label |
| Grilled chicken, rice, plain veg | Decent, as long as nothing was cooked in much oil |
| Restaurant plates, fried food, saucy dishes | Often hundreds of calories low |
| Mixed home cooking, stews, casseroles | Weak. The camera can't see the butter or oil |
So if your diet is mostly whole foods you can see, a photo app will get you in the neighborhood. If you eat out, order in, or cook with real fat, the photo is where the missing calories hide. My longer take on this is in voice tracking vs photo apps.
What Cal AI says about its own accuracy
To be fair to Cal AI, it does not claim to be perfect. Its own FAQ puts the app at about 80 percent accurate and adds that no food tracking app is. Its marketing has quoted 90 percent on visible foods, and that visible part is the whole point: on foods where the calories are visible, it does well. The trouble is that a lot of real eating is not visible.
Eighty percent accurate also sounds better than it feels. If your day is 2,200 calories and the count runs 20 percent low, that is about 440 calories a day you never logged. Over a week that is enough to cancel a careful deficit without you ever seeing where it went.
The fix: tell the app what the camera can't see
There is a simple way to close most of the gap, and it works no matter which app you use: say the part a photo can't. The oil, the butter, the dressing, the sauce.
That is the whole idea behind Logma, the app I build. Instead of snapping a picture, you say or type one sentence and it counts every item and logs the calories and macros in about 3 seconds. Because you are describing the food, you can include the stuff a camera would miss.
"Grilled chicken salad with two tablespoons of ranch and a handful of croutons."
That sentence carries the ranch, which a photo would have dropped. You can say weights when you know them, like "6 ounces of chicken," and skip them when you don't. It works in English and Turkish, there's no account to create, and your meals stay on your phone. Logma is iPhone only and needs iOS 17 or later, and it costs $49.99 a year after a 3 day free trial, with voice included rather than sold as an add on.
Cal AI vs talking: which one fits you
This is not really Cal AI versus Logma. It is photo versus describing, and the right pick depends on what you eat and how you like to log.
- Pick a photo app like Cal AI if you eat a lot of simple, visible foods, you like the tap and snap habit, and you would rather not talk to your phone. It is fast and it is fun, and for a plate of plain food it is close.
- Pick a voice or text app like Logma if you eat out often, cook with oil and butter, or want to log the sauce and the second helping without fighting a search bar.
- Pick a search and scan app like MyFitnessPal if most of what you eat has a barcode. A label beats any guess, photo or voice.
If you want the full field side by side, I compared the main ones in the best calorie tracking apps of 2026, and I wrote a separate piece on Cal AI alternatives if you have already decided to move on.
How to get a more accurate number from any photo app
If you are staying with Cal AI or another photo app, a few habits pull the number back toward reality:
- Add the cooking fat by hand. After the photo, log the oil or butter the dish was cooked in. One tablespoon of oil is about 120 calories.
- Log sauces and dressings separately. These are the single biggest hidden source on restaurant plates.
- Scan, don't shoot, anything with a label. A barcode is exact. A photo of a protein bar is a guess.
- Weigh the two or three foods you eat most. You don't need to weigh everything, just the staples. More on that in tracking without weighing every day.
- Trust your weekly trend, not one day. If the scale isn't moving the way the app says it should, the app is probably running low. Adjust your target with a fresh number from the free calorie calculator.
The point is not to abandon tracking. It is to stop letting the easy 300 calories slip through.
FAQ
Is Cal AI accurate?
For simple, visible foods it is reasonably accurate, and Cal AI's own FAQ puts it at about 80 percent. For restaurant plates, fried food and mixed dishes it tends to run low, because a photo can't see the oil, butter and sauce. A 2026 NIH study found photo apps, Cal AI included, undercounted real meals by 250 to 345 calories on average.
How much does Cal AI underestimate calories?
In a 2026 study from the National Institutes of Health, four photo apps including Cal AI underestimated meals by about 250 to 345 calories each, roughly a third of the real total, and underestimated fat by around 30 grams per meal. The study was presented at a conference and is still preliminary.
Why do photo calorie apps underestimate fat?
Fat carries about 9 calories a gram and a lot of it is invisible: the oil in the pan, the butter on the steak, the dressing in the salad. A camera reads shape and size, not what soaked in, so the hidden fat gets left out. Thirty grams of missed fat is about 270 calories.
Is Cal AI or MyFitnessPal more accurate?
In the NIH testing both ran low by a similar amount, because both relied on the same photo input. MyFitnessPal is more accurate when you scan a barcode or pick a verified entry, since a label is exact. For home cooked and restaurant food, describing the meal in words tends to beat a photo in either app.
Can I make Cal AI more accurate?
Yes. Add the cooking oil or butter by hand after the photo, log sauces and dressings separately, scan barcodes instead of photographing labeled foods, and trust your weekly weight trend over any single day's count. Those few habits pull the number much closer to reality.
Does Logma use photos?
No. With Logma you say or type what you ate in a sentence and it logs the calories and macros in about 3 seconds. Because you describe the food, you can include the oil, butter and sauce a camera would miss. It's iPhone only, works in English and Turkish, and needs no account.
Sources: American Society for Nutrition / NIDDK, Photo-based calorie-tracking apps may underestimate energy in meals (EurekAlert, July 2026) · ScienceDaily, Your AI calorie-tracking app may be off by 345 calories (July 2026) · U.S. News, How Accurate Are Photo-Based Calorie Apps? 4 Are Put To The Test (July 2026) · Cal AI official FAQ · FDA, How to Understand and Use the Nutrition Facts Label (calories per gram). Numbers are approximate and can change; nothing here is medical advice.
