Are AI Calorie Counting Apps Accurate? What Studies Show

Snap a photo, get the calories. AI calorie counters make logging fast, and that’s a real advantage. But how close is the number on the screen to what’s actually on your plate? Researchers have started to test exactly that, and the answer is useful once you know where the errors come from.

The short answer: AI calorie apps are good at recognizing what you’re eating but much weaker at judging how much. In controlled studies, general AI models estimated the calories in a meal photo with an average error of about 35%, and they tend to underestimate large portions and fatty foods. Correcting the portion and adding hidden oil or sauce closes most of that gap.

What the studies found

Here are the most relevant tests so far. They used different methods, so compare the patterns rather than the exact figures.

Study	What was tested	Main result
Fridolfsson et al., 2025	ChatGPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro on 52 weighed food photos	Average calorie error of 35.8% for ChatGPT and Claude, 64.2% for Gemini. All underestimated more as portions grew.
O’Hara et al., 2025	ChatGPT-4 on 114 meal photos from a national diet survey	Foods correctly identified 93% of the time. Portion weight off by 27.8% on average, with large meals estimated at about 530 g instead of about 800 g.
NIH test presented at NUTRITION 2026 (preliminary)	Four photo apps on 102 meals prepared in a metabolic kitchen	Calories underestimated by about 250 to 345 kcal per meal on average, fat by about 30 g. High fat meals were the hardest.
Martin et al., 2012	Photos analyzed by trained staff (Remote Food Photography Method)	Lab meals within about 1% of weighed values; daily intake within about 4% of measured energy needs.

Two things stand out. First, the photo itself isn’t the problem: when trained people analyzed meal photos carefully, the method was very accurate. Second, fully automatic estimates lose accuracy exactly where portions get bigger and fat gets hidden.

One detail from the ChatGPT study is worth knowing. Averaged over all 114 meals, the total energy estimate was almost spot on (0.1% difference), because overestimates and underestimates cancelled out. For any single meal, though, the error could be large. Over weeks of logging, errors that go in both directions partly average out. Errors that always go the same way (like missing oil) don’t.

Why AI misjudges portions

A photo is flat. The model has to guess depth, density and weight from a 2D image, and several things make that hard:

• Volume. A thick steak and a thin one can look the same from above. A deep bowl of pasta looks like a shallow one.
• Hidden fat. Oil in the pan, butter in the rice, dressing soaked into a salad. Fat is the most energy dense nutrient (9 kcal per gram) and often invisible.
• Mixed dishes. Lasagna, curries, stews and smoothies hide their ingredients, so the model has to assume a recipe.
• Big plates. Every study above found that underestimation grows with portion size. Small, simple portions were estimated much better.

What AI does well

It’s not all bad news. AI apps are strong at:

• Recognizing foods. 93% precision in the ChatGPT study.
• Simple, single items. A banana, an apple or a slice of bread are easy to identify and have standard sizes. (If you’re curious, here’s how many calories are in a banana by size.)
• Ranking meals. Even when the exact number was off, the estimates correlated well with the real values (0.73 for energy in the ChatGPT study), so the app still tells a heavy meal from a light one.
• Speed. Logging in seconds means you actually do it, every day.

Is it worse than counting calories by hand?

Not really. Counting by hand has its own blind spots. In a classic New England Journal of Medicine study, people who couldn’t lose weight on a diet they reported as under 1,200 kcal were actually eating much more: they underreported their intake by about 47%, while their metabolism was normal.

The Swedish researchers who tested three AI models concluded that the best ones reached accuracy comparable with traditional self-reporting, without the effort. That’s a fair summary: AI is not a lab scale, but neither is your memory.

A worked example

Say your real intake is 2,100 kcal a day and your app is consistently low by 30% on dinner only:

Meal	Real calories	Logged by photo
Breakfast (yogurt, oats, fruit)	450 kcal	440 kcal
Lunch (sandwich, apple)	650 kcal	610 kcal
Dinner (pasta with sauce and oil)	1,000 kcal	700 kcal
Total	2,100 kcal	1,750 kcal

The gap is 350 kcal a day, which is most of a typical weight loss deficit. Now suppose you add “1 tablespoon olive oil” and bump the pasta to the amount you really served. Dinner moves close to 1,000 kcal and the log becomes useful again. The fix took ten seconds.

How to get accurate results from a photo app

1. Shoot from above, whole plate in frame, with a fork or your hand nearby for scale.
2. Add what the camera can’t see: cooking oil, butter, dressing, sugar in coffee.
3. Check the portion on big or mixed plates and correct it if it looks low.
4. Scan the barcode for packaged food instead of photographing it: the label is more reliable than any estimate.
5. Weigh your staples for a week. Pasta, rice, cereal and nuts. Once your eye knows what 80 g of dry pasta looks like, your corrections get better.
6. Trust the trend, not a single day. If your weight doesn’t move after two or three weeks, lower your target a little rather than doubting every entry.

In BananaCal you can do all of this in one place: snap a photo, scan a barcode, or just type or say “plus a tablespoon of oil” and the estimate updates.

Does perfect accuracy even matter?

Less than you’d think. A systematic review of weight loss studies found a consistent link between keeping a food record and losing weight. The benefit comes from doing it regularly, not from getting every gram right. Pick a method that’s fast enough to keep using, fix the big systematic errors (oil, portions, drinks), and let your weight trend tell you whether your target is right.

If you haven’t set that target yet, use our calorie calculator and read how many calories to eat to lose weight. Tracking protein alongside calories helps too, and AI apps tend to estimate protein better than fat.

Quick checklist

• AI apps recognize food well but underestimate portions, especially large and fatty ones.
• Expect an error of about a third on a single complex meal if you don’t correct anything.
• Always add oil, butter, sauces and drinks by hand.
• Use the barcode for packaged food.
• Consistency beats precision: log every day and adjust by your weight trend.