AI Calorie Apps Undercount by Up to 345 Calories — Oxalate Is Harder to Photograph Still

2026-08-31 · By Elena Novak, Research Editor · Dietary reference, not medical advice

Photo-based nutrition tracking has quietly become the default: point a phone at a plate, let the model name the foods, accept the number. A study out of the NIH Clinical Center presented at Nutrition 2026 — the American Society for Nutrition's annual meeting — put that convenience to the most controlled test it has had, and the result is a useful warning for anyone who assumes the app's estimate is the meal.

The finding, plainly: researchers at NIDDK (Aaron Hengist, Olivia Charles and colleagues) photographed 102 meals prepared in a metabolic kitchen where every ingredient was weighed to the nearest 0.1 gram, then ran the photos through four popular apps: MyFitnessPal, LoseIt!, CalAI and Appediet. All four underestimated. Calories came back 250 to 345 kcal too low per meal on average, and fat was short by roughly 30 grams — about a third of the true load. Carbohydrate estimates were the most consistent; high-fat meals, especially ketogenic ones, produced the worst errors. A follow-up run of 200-plus meals pointed the same direction. One caveat the researchers state themselves: this is a conference presentation, reviewed but not yet peer-reviewed, so the numbers are preliminary.

The Direction of the Error Is the Story

A random error averages out over a week; a one-directional error compounds. Someone logging three meals a day through a photo could undercount by 750 to 1,000 kcal daily without any signal that the estimate is systematically low — which is exactly the failure mode that makes people conclude their diet "isn't working." The mechanism is visible in the study design: a camera cannot see oil absorbed into a stir-fry, butter under a sauce, or the difference between two cuts of meat. The nutrients that hide inside appearance are the ones that get missed.

Oxalate: The Nutrient No Camera Can Estimate

Calories at least have visual proxies — portion size, browning, recognizable dishes. Oxalate has none. A spinach smoothie and a kale smoothie look identical through a lens; one carries 493 mg of oxalate per cooked half-cup equivalent and the other 3 mg. Almond flour pancakes photograph like any pancake while carrying 197 mg per quarter cup of flour. And the app database layer is thinner still: oxalate values are sparse, inconsistent about serving sizes, and rarely lab-measured — the exact field where our ion-chromatography dataset does its work.

Food	Oxalate (mg)	Serving	Rating
Spinach, boiled or steamed	493	1/2 cup	Avoid
Almond meal / almond flour	197	1/4 cup	Avoid
Almonds, raw or dry roasted	107	1/4 cup	Avoid
Dark chocolate, bar or chips	65	1 oz	Caution
Peanut butter, creamy or chunky	36	2 tbsp	Caution

Every food in that table is one a photo-based tracker would happily log as "salad," "pancakes," "snack," or "smoothie" with no oxalate field at all. The failure is not sloppiness — it is category error. Oxalate is a composition property measured by ion chromatography in a lab, not a visible feature a vision model can infer.

The Practical Split

None of this says throw the apps away; it says give each tool the job it can do. Photo apps handle the energy arithmetic — with the study's caveat that users should adjust portions manually rather than accept the auto-estimate, and remember fat is where the error concentrates. Oxalate management needs the other pipeline: curated, lab-measured values tied to honest serving sizes, the pairing-with-calcium rule, and the 24-hour urine test that tells you whether any of the arithmetic is moving your actual risk. The NIH study is, at bottom, a defense of measurement — which is a position this site holds for every nutrient, not just calories.

Frequently Asked Questions

Which apps were tested, and how big was the error?

MyFitnessPal, LoseIt!, CalAI and Appediet, tested against 102 metabolic-kitchen meals weighed to 0.1 g. Calories ran 250-345 kcal low per meal on average and fat about 30 g low; carbohydrate estimates were the most reliable. Findings were presented at Nutrition 2026 and are preliminary pending peer review.

Can AI apps estimate oxalate if I photograph a food?

Not credibly. Oxalate is invisible in appearance — spinach and kale smoothies look alike but differ by roughly 490 mg per cooked half-cup — and most nutrition databases carry sparse, non-standardized oxalate entries. Lab-measured values with defined serving sizes are the only dependable source.

What should I use instead for oxalate tracking?

A curated database with measured values (like this site's), attention to serving size, calcium-with-meals pairing for the oxalate you do eat, and — for stone formers — a 24-hour urine collection, which measures what your kidneys actually excreted rather than what a camera guessed.

Sources:
Harvard T.H. Chan School of Public Health Oxalate Database (2024); USDA FoodData Central; American Society for Nutrition via ScienceDaily: Your AI calorie-tracking app may be off by 345 calories (July 26, 2026).; HealthDay coverage, July 27, 2026. All oxalate values are lab-measured via ion chromatography. This page is a dietary reference tool, not medical advice.

See how oxalate values are measured and verified on our data methodology page.

Medical Disclaimer: This site is a dietary reference tool based on published food composition data. It does not provide medical advice, diagnose, or treat any condition. Talk to your doctor or a registered dietitian before changing your diet, especially if you have kidney stones, kidney disease, or other medical conditions.