Snap a photo, type it in, scan the barcode or weigh it on a kitchen scale: people argue about which of these is the accurate way to log food. In the research, every method undercounts somewhere, including the one scientists use as their reference. Knowing where each one fails is more useful than picking a winner.
Everyone underreports, and some people a lot
The best test of a food log is to compare it with what the body actually burned, measured with doubly labeled water, a tracer method that works out energy expenditure over one to two weeks. In a stable-weight adult, intake and expenditure should roughly match.
In the OPEN study, 484 adults aged 40 to 69 did exactly that comparison (Subar et al., 2003). On 24-hour recalls, where a trained interviewer walks you through yesterday’s food, men underreported energy by 12 to 14% and women by 16 to 20%. On food frequency questionnaires the gap was about a third.
The extreme case is a 1992 study in the New England Journal of Medicine (Lichtman et al.). Ten people who said they could not lose weight on under 1,200 kcal a day were measured carefully for 14 days. Their metabolism was normal for their body composition. They had underreported what they ate by an average of 47% and overreported their exercise by 51%. The authors found no distinct psychological problems in the group, and the participants described their own eating as fairly normal. Forgotten snacks and portions rounded down add up over a week.
Portions are where the numbers drift
Even when you remember every item, you have to say how much. In a 2026 study in the Journal of Nutrition, 128 adults estimated the weight of 59 foods and drinks from photos (Nurmanova et al.). Without help, their estimates were off by about 80% on average (mean absolute percentage error). With a photo atlas of reference portions, the error dropped to about 45%. An AI model trained on the same cuisine landed in between, at about 68%, and struggled most with small portions and visually complex dishes.
The errors that hurt most in a calorie log are on dense foods. A serving of granola in USDA’s data is 50 g, about 210 kcal; a free pour into a big bowl is easily 80 g, about 335 kcal. Two tablespoons of peanut butter are 32 g and about 190 kcal, and a heaped knife from the jar rarely matches that.
Weighing is the reference, and it still slips
Weighed food records are what researchers fall back on when they need precise intake from free-living people. A 2015 review describes them as often treated as a reference method, while noting the high burden on participants, records getting worse the more days you ask for, and people changing what they eat to make recording easier (Ortega et al.).
When a team in Northern Ireland checked seven-day weighed records against doubly labeled water, the records averaged 9.66 MJ a day (about 2,310 kcal) against a measured expenditure of 12.15 MJ (about 2,900 kcal), roughly 20% low (Livingstone et al., 1990). People in the top third of recorded intakes were accurate. In the lower two thirds, three of the four subgroups recorded only 61 to 74% of what they burned.
Photos, barcodes and typing
Photo apps are the fastest and the least precise on fat. We covered the 2026 NIH test of four photo apps in detail in Are AI calorie apps accurate?. The short version is that the apps came in about a third low on calories, mostly because oil and butter are invisible in a picture.
Barcode scanning is close to the opposite. In a 2025 trial, 44 adults were given preweighed food for three days and logged it in MyFitnessPal, which 86% of them could do by scanning barcodes (Diktas et al.). On average the app was only 3.7% (76 kcal a day) away from the weighed truth, though the spread between people was wide (a standard deviation of 316 kcal a day). Two catches remain. The label itself can be off: US rules allow real calories up to 20% above the stated number, which we explain in why your calorie number should be a range. And user-submitted database entries can be wrong; a 2025 study blamed MyFitnessPal’s poor validity for athletes’ food records partly on non-verified consumer entries, while Cronometer did much better (Morello et al.).
Typing or saying what you ate sits between the two. It depends on your memory and your portion guesses, so it inherits the recall errors above, but you can add detail a photo cannot show, such as “cooked in two tablespoons of oil” or “only half”.
How the methods compare
| Method | Typical accuracy | Effort | Where it fails |
|---|---|---|---|
| Weighing everything | Best for what gets logged; whole records still ran about 20% low against measured burn | High, every meal | Skipped snacks and drinks, eating out, people eating differently to make logging easier |
| Barcode or label | About 4% average error with packaged food in one trial | Low for packaged food | Label tolerance, wrong database entries, home cooking and restaurants |
| Typing or talking | As good as your memory and portion guesses | Low | Forgotten items, portions rounded down, hidden fat |
| Photo | About a third low on calories in the NIH test | Lowest | Oil, butter and sauces, bowl depth, small portions |
A hybrid that works
You do not need to weigh food forever. This is the setup we suggest, based on the evidence above:
- Weigh your regular foods a few times. The breakfast cereal, the pasta, the peanut butter, the oil in the pan. One week is usually enough to learn what your bowl and your spoon actually hold. Coach Eric Helms suggests a similar self-check: estimate a day by eye, then count it carefully and compare.
- Then log by photo or text, and correct it. Add the oil, the dressing and the second helping in words, and fix the portion when you know it.
- Scan barcodes for packaged food, and enter the amount you actually ate, not the serving on the label.
- Watch the 7-day average weight over two or three weeks. If it is not moving the way your TDEE calculator said it would, the log is probably low. Lower your target by 100 to 200 kcal and give it another two weeks.
ChatMyFood is built for steps 2 and 4. You log by photo or by typing, every estimate comes as a range, and it asks a short question when the oil or the portion would change the number a lot. It is free to try, and it works best alongside a few days with a kitchen scale.