AI Diet-Tracking Apps: Expert Test Reveals Surprising Risks

AI diet-tracking apps promise to turn the chore of logging every meal into a photo, a barcode scan or a spoken sentence. Sibylle Kranz, a registered dietitian and associate professor at the University of Virginia School of Education and Human Development, wanted to know whether the promise holds. She signed up for free trials of several popular apps, built different user profiles with different health goals, and compared the features and the advice each one gave her. UVA Today published her verdict on 31 August 2026, and it is a mixed bag: real convenience, real personalisation, and accuracy that she describes as “a guessing game still”.

This article sets out what Kranz found, why her hands-on experience lines up with three recent peer-reviewed and conference studies of food image recognition, and what anyone thinking about an AI diet-tracking app should check before trusting it. It follows our earlier review of CalBye, an AI nutrition app, and our recent piece asking what is left for human doctors as AI advances. The same tension runs through all three: software that is good enough to be useful and not yet good enough to be trusted on its own, whether the question is accuracy or data protection.

Table of contents

Who Tested the AI Diet-Tracking Apps and How
What the AI Diet-Tracking App Gets Right
Where the AI Diet-Tracking App Falls Down: Accuracy
The Studies That Back Her Up
AI Diet-Tracking App Privacy: The Cost of Personalisation
How to Test an AI Diet-Tracking App Yourself
Supplement, Not Replacement: The Dietitian’s Verdict
What Better AI Diet-Tracking Apps Might Look Like
What This Means for Users and Developers
Conclusion
References and Further Reading

Who Tested the AI Diet-Tracking Apps and How

Kranz is not a casual reviewer. She directs the Diet and Nutrition laboratory at UVA, studies how AI, mobile apps and wearable devices support healthier eating habits, and was recently named an Excellence in Nutrition Fellow of the American Society for Nutrition for her contributions to nutrition science and public understanding. In the UVA Today report by Laura Hoxworth, she describes an informal but structured test rather than a formal trial.

The method behind the test

Kranz created several user profiles, each with a different health goal, and ran them through free trials of the most popular AI diet-tracking apps on the market. UVA Today does not name the AI diet-tracking apps she used, and it gives no numeric error figures, so her findings are qualitative. What she measured instead was the experience a real user has: how the app interprets a photo, what happens when the same meal is logged twice, what data it asks for, and what it does with the answers.

Why an informal test still matters

A dietitian running an AI diet-tracking app through its paces catches things a controlled laboratory study cannot. Laboratory work, as we show below, photographs meals under good lighting at a fixed angle. Kranz photographed what she actually ate, over weeks, and watched the app’s behaviour drift. Both kinds of evidence are needed, and the interesting result is that they agree.

What the AI Diet-Tracking App Gets Right

Kranz starts from the observation that accurately tracking what you eat is notoriously difficult, and that generative AI has transformed the category. Diet-tracking apps have existed for decades, but the AI diet-tracking app is a different product, but the manual meal log has been replaced by a barcode scan, a spoken description or a photograph that returns an instant analysis of portion size and nutrients. Some apps add custom meal plans and nutrition coaching on top.

Convenience that keeps people consistent

Compared with a handwritten food diary or regular appointments with a dietitian, Kranz said, an AI diet-tracking app makes meal tracking easier and more accessible. That matters because consistency is the single biggest predictor of whether tracking works at all. A tool that removes friction helps people stay motivated and reach their goals, even if its individual estimates are imperfect.

Personalisation a generic plan cannot match

As the apps learn a user’s habits, they recognise subtle patterns and tailor recommendations in ways that a generic meal plan, or even a professional dietitian working from fortnightly appointments, cannot. “Over time, it’s almost as if the AI knows what you have in your pantry and in your fridge,” Kranz said. “It knows things about you that you don’t know about yourself.”

Honesty that is easier with software

Kranz also noted that eating and drinking are intensely personal, and that someone uncomfortable sharing intimate details with a clinician may be willing to share them with an app. An AI diet-tracking app can therefore capture a more honest record than a face-to-face consultation. The table below summarises the upside she reported.

Feature	What Kranz found	Her verdict
Photo, barcode and voice logging	Replaces the manual food diary; instant portion and nutrient estimate	Easier and more accessible than paper or appointments
Pattern recognition	Learns pantry contents and habits over time	Personalisation beyond a generic plan
Private disclosure	People share with an app what they hide from a clinician	More honest records
Photo recognition accuracy	“Dozens of examples” where a photo had to be corrected	“A guessing game still”
Crowdsourced databases	Same meal identified differently a week later	“The accuracy is just not there”
Data collection	Eating habits, location, body measurements, emotional well-being	“A big privacy risk”
Business model	Designed to keep you using and to sell a subscription	Keep “a healthy dose of skepticism”

Where the AI Diet-Tracking App Falls Down: Accuracy

The central finding is about accuracy. Kranz said recent studies have found that AI sometimes underestimates calorie counts or provides incorrect dietary advice, and that her own testing backed those findings up. “I could give you dozens of examples where I took a photo of something I ate, and then I had to correct it because it misinterpreted what’s there,” she said. “Although AI has come a long way, it’s a guessing game still.”

Misread photographs

The failure mode Kranz describes is not a rounding error. The AI diet-tracking app misidentified what was on the plate, so the calorie and nutrient figures that followed were built on the wrong food. A user who does not check the recognition step inherits every downstream mistake, and most users do not check, because not checking is the point of using a photo instead of a form.

Crowdsourced databases that shift under you

Some apps rely on crowdsourced databases built from user-submitted entries, which creates inconsistent or inaccurate information, and as more users add more data the recommendations keep shifting. Kranz saw this directly. “I took a photo of something it recognized, confirmed it, and put it in the database. A week later, I had the exact same thing, and it identified it differently or changed some of the ingredients,” she said. “The accuracy is just not there.”

Why this matters for anyone in a calorie deficit

A misidentified ingredient or a shifting database entry is a nuisance for someone tracking patterns. It is a real problem for someone managing a medical diet or trying to hit a specific energy target, because an AI diet-tracking app that quietly under-reads a meal by a few hundred calories can turn an intended deficit into maintenance without the user ever knowing.

The Studies That Back Her Up

Kranz referred to “recent studies” without naming them. Three stand out, and between them they cover controlled meals, everyday photographs and a proposed fix. The table below sets them side by side before we look at each.

Study	Who and when	Method	Headline result
NIH photo-app test	NIDDK, NUTRITION 2026, July 2026 (preliminary)	102 metabolic-kitchen meals photographed and uploaded to MyFitnessPal, Lose It!, CalAI and Appediet	Calories underestimated by 250 to 345 per meal; fat by about 30 g
University of Sydney app review	Li et al., Nutrients, August 2024 (peer-reviewed)	800 apps screened, 18 quality-scored, 7 AI photo-recognition apps tested on 22 images with 39 food components	Recognition accuracy from 46% to 97% depending on the app
Emory NutriCamp	Yan and Luo, Communications Medicine, August 2026	New photo, text and voice app benchmarked on everyday multi-food meals	63% lower average error than competing apps and GPT Vision
UVA hands-on test	Kranz, UVA Today, 31 August 2026	Free trials of several popular apps under multiple user profiles	Repeated misreads, shifting database entries, privacy concerns

The NIH metabolic-kitchen test

The most rigorous AI diet-tracking app accuracy test to date came from the National Institute of Diabetes and Digestive and Kidney Diseases. Postdoctoral fellow Aaron Hengist and research fellow Olivia Charles used an ongoing study at the NIH Clinical Center where every meal is prepared in a metabolic kitchen with ingredients weighed to the nearest 0.1 gram, so the true calorie content of each plate was known before it was photographed. They photographed 102 meals and uploaded the images to four popular apps.

The results, presented at NUTRITION 2026 in July, showed the apps’ calorie totals were about 250 to 345 calories too low per meal on average, and fat was underestimated by roughly 30 grams. High-fat ketogenic dishes caused the most trouble, carbohydrates were estimated most consistently, and MyFitnessPal and Lose It! did better on higher-calorie meals than on lower-calorie ones. “People using a photo-based tracking app without adjusting the portions or entering the amounts of food should take the results with a grain of salt,” Hengist said. The American Society for Nutrition notes the abstract has not completed journal peer review and should be treated as preliminary.

NIH test: average calorie shortfall per meal across four AI diet-tracking apps (kcal, low and high end of the reported range)

High end of range 345 kcal

Low end of range 250 kcal

Fat shortfall, in kcal terms (30 g at 9 kcal per gram) 270 kcal

The University of Sydney app review

The peer-reviewed evidence comes from Xinyi Li, Juliana Chen and colleagues at the University of Sydney, published in Nutrients in August 2024. They screened 800 apps from the Australian Apple and Google stores, scored 18 for quality and behaviour-change potential, and tested seven with AI photo recognition on 22 images containing 39 food components. Accuracy was simply the share of components the app identified correctly.

The spread was enormous. MyFitnessPal identified 38 of 39 components, or 97%; Fastic reached 92%, HealthifyMe 90% and Foodvisor 87%. HitMeal managed 62%, and Lose It! and FatSecret each identified only 18 of 39, or 46%. Recognising the food is only the first step, though: on single items MyFitnessPal’s energy estimate was within 3% of the truth, HealthifyMe over-read by 8%, while Fastic over-read by 44% and Foodvisor under-read by 47%.

Food component recognition accuracy in the Sydney test (share of 39 components identified correctly)

MyFitnessPal 97%

Fastic 92%

HealthifyMe 90%

Foodvisor 87%

HitMeal 62%

Lose It! 46%

FatSecret 46%

Manual logging is not safe either

The Sydney team also checked the older, manual side of each AI diet-tracking app by logging three set diets by hand in 16 apps. On a Western diet, 14 of the 16 apps overestimated energy, by 1,040 kJ on average; on an Asian diet, 13 of 16 underestimated, by 1,520 kJ; and on a recommended healthy diet every single app underestimated, by 944 kJ. The database problem Kranz describes is older than the AI layer that now sits on top of it.

Average energy error from manual logging across 16 apps in the Sydney test (kJ, size of error)

Asian diet, underestimated 1,520 kJ

Western diet, overestimated 1,040 kJ

Recommended diet, underestimated 944 kJ

AI Diet-Tracking App Privacy: The Cost of Personalisation

The second half of Kranz’s verdict concerns data. To generate custom recommendations, an AI diet-tracking app collects extensive personal information: eating habits and location data, body measurements and even emotional well-being. “Unless the user carefully reads the data protection policies, they don’t really know what the company is doing with all this data,” she said. “That’s a big privacy risk.”

What the app knows about you

The same learning that lets an AI diet-tracking app guess your fridge contents also builds a detailed behavioural profile. Location data shows where you eat; timestamps show when; mood check-ins show why. Combined with weight, height and health goals, that is sensitive data under most privacy regimes, and Kranz’s point is that few users read the policy that governs it.

Choosing an AI diet-tracking app that lets you say less

Her advice is practical. Review the privacy settings carefully and choose an AI diet-tracking app that lets you limit the personal information you share. If the goal is to identify eating patterns, you do not need to provide every detail the app asks for, and a well-designed one will still work with less.

How to Test an AI Diet-Tracking App Yourself

Kranz’s recommendations amount to a six-step evaluation of any AI diet-tracking app that anyone can run during a free trial. We have laid them out as a checklist, with the reason each step matters drawn from her comments and from the studies above.

Step	What to do	Why it matters
1. Read the privacy settings	Limit sharing to what your goal actually needs	Apps collect location, body and mood data by default
2. Verify accuracy	Compare the app’s figures with a trusted nutrition source	NIH found 250 to 345 kcal per-meal shortfalls
3. Watch for follow-up questions	Prefer an app that asks before it assumes	Assumed portions drive the biggest errors
4. Commit to a full week	Track everything, “every piece of gum, every glass of water”	Drift and inconsistency only show over time
5. Judge the outcome	Ask how happy you are with what it told you	Usefulness, not novelty, is the test
6. Stay skeptical	Remember the app wants continued use and a subscription	Recommendations may serve retention as much as health

Verify against a trusted source

Kranz recommends verifying the app’s accuracy by comparing its recommendations with trusted nutrition sources. In practice that means weighing a few staple foods once and checking whether the AI diet-tracking app’s photo estimate lands near the label value, exactly as the NIH team did with a metabolic kitchen. If it consistently under-reads fat, as the NIH apps did by around 30 grams, you know which meals to correct by hand.

Give it a week and track everything

She also recommends paying attention to whether the app asks follow-up questions before making assumptions, and giving it enough time to prove itself by committing to a weeklong test. “Be 100% committed and track everything, every piece of gum, every glass of water,” she said. “And then look at how happy you are with the outcome.”

Remember who the app works for

Her closing advice is the one most reviews leave out. Exploring an AI diet-tracking app as a helpful assistant is a good idea, Kranz said, but keep a healthy dose of skepticism. “At the end of the day, they want you to continue using the app,” she said. “Better yet, they want you to buy the subscription.”

Supplement, Not Replacement: The Dietitian's Verdict

Despite the limitations, Kranz believes an AI diet-tracking app can be a valuable supplement to professional nutritional guidance, although not a replacement. “You have the diagnosis, you have the expert evaluation, and then you use this tool to support that,” she said. The order matters: the clinician sets the target, the software helps the patient hit it.

Where the app belongs in a care plan

That framing places the AI diet-tracking app in the same position as a home blood-pressure cuff or a step counter: a device that produces useful, imperfect data between appointments, to be interpreted by someone who knows its error bars. It is the same conclusion we reached when we asked whether AI leaves anything for human doctors to do, and the answer there, as here, was judgment.

Where it does not belong

An AI diet-tracking app is not a substitute for a diagnosis, and it is a poor tool for anyone whose health depends on hitting an exact number, whether a renal patient tracking potassium or an athlete cutting weight. For those users the NIH finding, that a photo can under-read a meal by 250 to 345 calories, is disqualifying without manual correction.

What Better AI Diet-Tracking Apps Might Look Like

The accuracy gap is not permanent. Two weeks before the UVA piece, Emory University announced NutriCamp, a dietary assessment app built by Runze Yan of the Nell Hodgson Woodruff School of Nursing and Hanqi Luo of the Rollins School of Public Health. Users record meals by photo, text or voice, and the system estimates 65 nutrients and food components from a database of 5,624 foods and beverages and more than 23,000 portion descriptions.

Cutting the error by 63%

In a study published in Communications Medicine, NutriCamp reduced average estimation error by 63% for food weight and four key nutritional measures compared with three popular AI diet-tracking apps, existing computer vision models and GPT Vision, when tested on photographs of everyday meals containing multiple foods. Yan’s framing echoes Kranz: “NutriCamp is clinically relevant because most existing apps focus on lifestyle, appearance and calorie counting.”

What the next generation of AI diet-tracking apps needs

Read together, the evidence points to three fixes. First, a curated nutrient database rather than a crowdsourced one, so a confirmed meal does not change identity a week later. Second, explicit portion questions before an estimate is committed, which is the follow-up behaviour Kranz looks for. Third, published validation against weighed meals, so users can see an app’s error bars before they subscribe. Building that kind of model is a serious machine learning model development exercise rather than a wrapper around a general-purpose vision model, and it is where the category is heading.

What This Means for Users and Developers

For users, Kranz’s test and the studies behind it lead to a simple rule: use an AI diet-tracking app for patterns and motivation, correct its portions by hand, and never let it be the only source of truth for a medical diet. For developers, the market is telling them that recognition accuracy in a controlled photo is no longer the differentiator; database stability, portion prompting and transparent validation are.

For businesses building in this space

Anyone building a health or fitness product should treat the NIH and Sydney figures as the bar to beat and publish their own validation. Our guide to building a fitness app successfully covers the product side, and the wider trend of on-device intelligence in our piece on AI in mobile phones explains why the next AI diet-tracking app may do more of its recognition locally, which also answers part of Kranz’s privacy concern.

For everyone else

The honest summary of what a nutrition expert found is that the AI diet-tracking app is a good assistant and a poor authority. That is a fair description of most consumer AI in 2026, and it is worth keeping in mind the next time an app tells you, with great confidence, exactly how many calories were on your plate. You can follow the wider category in our AI models and tools hub.

Conclusion

Sibylle Kranz set out to see whether an AI diet-tracking app could do the job of a food diary and a dietitian, and came back with a clear split decision. On convenience, consistency, personalisation and honesty, the AI diet-tracking apps win. On accuracy they do not: photos are misread, crowdsourced databases shift, and independent studies put the typical per-meal calorie shortfall at 250 to 345 kcal. On privacy they demand more than most users realise they are giving.

Her advice, to review the settings, verify the numbers, insist on follow-up questions, commit to a full week and stay skeptical about who the AI diet-tracking app really serves, is the best short guide to using one well. Treat the AI diet-tracking app as a supplement to expert advice, and it earns its place; treat it as the expert, and it will let you down by a few hundred calories a meal.