Bo guide
Short answer
AI calorie apps land within 10–15% on single recognisable foods, but 25–35% off on mixed meals, and one 2026 study found four popular apps underestimated calories and fat by roughly a third.
General information, not medical advice. Written from published sources, which are listed at the end. It is not a substitute for a doctor, pharmacist or registered dietitian who knows your situation.
In this guide8 sections · 4 min
Honest answer: less accurate than the marketing implies, and accurate enough to be useful anyway — provided you know where the error comes from.
Testing in 2026 found four popular AI food apps underestimated calories and fat by about one third against carefully prepared meals. Broader analysis puts them within 10–15% on single, clearly visible foods, and 25–35% off on complex mixed dishes.
Compared to what, though?
A third sounds damning until you ask what the alternative actually achieves. The benchmark is not a perfect measurement — it is what you would otherwise do, which is estimate.
The reference point here is a 1992 study that measured real intake with doubly labelled water rather than trusting food diaries. Participants who believed they were eating about 1,028 calories a day were actually eating 2,081 — under-reporting by roughly 47%, while genuinely trying to be accurate.
So the comparison is roughly a third low for an AI photo app against roughly a half low for an honest human estimate. Neither is measurement. But "the app is wrong" is not an argument for going back to guessing, which is wrong by more. The accurate option remains weighing your food, and it always was.
Where the error actually comes from
The failure is not really image recognition. Modern models identify food well. The problem is that the highest-calorie components of a meal are frequently invisible in a photo:
- Cooking oil. A tablespoon is about 120 calories and leaves no visual trace once absorbed.
- Dressing and sauce. Ranch versus vinaigrette can be a 300-calorie swing on an identical-looking salad.
- Butter. Melted into vegetables or a steak, it simply is not visible.
- Hidden sides. Rice or potatoes under a piece of chicken.
- Portion depth. A photo is two-dimensional. Bowl depth is genuinely hard to infer.
Accuracy varies by cuisine
Training data is uneven. American, Italian and Asian dishes tend to land around 85–90% accuracy, while less-represented cuisines — Indian, Middle Eastern and many regional dishes — sit closer to 75–80%. High-fat ketogenic dishes appear to cause the most trouble; carbohydrates are estimated most consistently.
How much to trust the studies
Worth a caveat, since this article rests on them. The published tests tend to use small numbers of carefully prepared meals with known composition, which is the right way to get a clean answer but not the same as everyday eating. Apps also change frequently, so a figure measured on one version may not describe the next.
The direction of the finding is consistent across tests — these apps underestimate, and they underestimate most on mixed, oily and restaurant food. The precise size of the gap is less settled than a single headline number suggests.
Three habits that close most of the gap
- Include a size reference. A hand, a fork, a standard plate. This alone improves estimates meaningfully because scale is the hardest thing to infer.
- Separate components before photographing. Push the rice away from the curry. Overlapping food hides mass.
- Correct the estimate when you know better. If you cooked it, you know how much oil went in. Editing the number takes five seconds and is the most accurate data the app will ever get.
Applied together, these are reported to improve accuracy by roughly 10–15 percentage points — larger than the difference between most competing apps. There is more detail on photographing food well if you want the full set.
What a third actually looks like
Abstract percentages are easy to shrug at, so here is the same error in food:
| Meal | Typical photo estimate | Realistic actual | Missed |
|---|---|---|---|
| Chicken salad with dressing | ~350 cal | ~620 cal | Dressing, oil |
| Stir fry with rice | ~500 cal | ~780 cal | Cooking oil, rice under the topping |
| Steak and vegetables | ~600 cal | ~820 cal | Butter on both |
| Pasta, restaurant portion | ~550 cal | ~900 cal | Portion depth, oil, cheese |
Four meals a day at that gap is roughly a thousand calories unaccounted for — enough to turn a deficit into maintenance without anything appearing wrong in the log. This is the single commonest reason a diet that "should" be working is not.
What to look for in an app
- It asks follow-up questions. Oil, dressing and butter are where the calories hide, and the only way to capture them is to ask.
- It lets you edit the estimate without fighting you about it.
- It has a barcode scanner. For packaged food a label lookup is exact, and a photo never will be.
- It shows a trend, not just a daily total. The weekly average is the number that means anything.
Does the inaccuracy matter?
It depends what you are doing. For general awareness and building consistency, a 15% error is not a problem — the trend is still real and the habit of logging is doing most of the work.
For a precise cut or a medical protocol, photo estimates alone are not enough. Weighing food remains the accurate method, and it is the first thing to try if the scale has stopped moving. A reasonable middle path is to weigh for a week to calibrate your eye, then use photo logging for maintenance.
