# How Accurate Are AI Calorie Apps? Every Published Study, 2023–2026

> Every study we could find on how accurately AI estimates calories from food photos, in one table: the 2026 NIH test of four apps, the ChatGPT, Claude and Gemini studies, two systematic reviews, Middle Eastern food research and vendor claims. Typical error is 14% to 36%, and it leans low.

Canonical: https://www.mycallie.app/en/blog/ai-calorie-accuracy-research
Published: 2026-10-20 · Updated: 2026-10-20
Author: Md Sadab Wasim · Source: https://www.mycallie.app

## Key takeaways

- Across published studies from 2023 to 2026, AI estimates of calories from a food photo are typically 14% to 36% off, and they lean toward underestimating.
- In the largest app test, an NIH team found MyFitnessPal, Lose It!, Cal AI and Appediet underestimated 102 weighed meals by 252 to 345 kcal each, about a third, with fat about 30 g low. It is a conference abstract, not yet peer reviewed.
- General chatbots did no better from a photo alone: 35.8% error for GPT-4o and Claude 3.5 Sonnet, 64.2% for Gemini 1.5 Pro (2025), and 30.5% for ChatGPT-5 (2025).
- Correction is the biggest lever in the data. Giving ChatGPT-5 the ingredient list with amounts cut its error from 30.5% to 13.9% across 195 dishes.
- Error grows with large portions, mixed dishes and hidden fat. A 2026 preprint on 802 meals adds that a small average error can still hide wide misses on single meals.
- Most of the evidence is small: 15 to 195 meals per study, several conference abstracts or theses, and models that change every few months.

Callie is our app. It appears here once, in our own 20-meal test, in the last row of the table. Every other figure links to its source, checked on October 20, 2026.

## The 30-second answer

**AI photo calorie estimates are typically 14% to 36% off, and they lean low.** From a photo alone, misses sit near the top of that range. NIH researchers found four popular apps underestimated weighed meals by [252 to 345 kcal, about a third](https://www.eurekalert.org/news-releases/1136415) (2026), and chatbots missed by 30.5% to 35.8% in 2025 studies. Accuracy improves sharply when a person corrects the ingredients and portions: giving ChatGPT-5 the ingredient list cut its error to 13.9%.

## Every study at a glance

Every study we found on AI calorie estimation from food photos, grouped by type: app and model tests, reviews, recognition studies, then vendor claims and our own test. Errors are for calories unless noted.

| Study and year | Journal or venue | What was tested | Sample | Reference method | Headline error | Notes |
|---|---|---|---|---|---|---|
| Charles, Hengist et al. (NIDDK), 2026 | NUTRITION 2026 abstract, American Society for Nutrition | MyFitnessPal, Lose It!, Cal AI, Appediet | 102 meals, plus 200+ in a follow-up | Metabolic kitchen, ingredients weighed to 0.1 g | Under by 252 to 345 kcal per meal (about one-third); fat under by about 30 g | Not peer reviewed. Keto meals did worse |
| Li, 2026 | arXiv preprint | 6 Claude and GPT models | 802 meals, 148 feeding days | Weighed intake in a controlled feeding trial | 3 of 6 models kept average error within 10%; none met the 30% per-meal limit | Not peer reviewed. Every model underestimated protein |
| Isobe et al., 2026 | Nutrients 18(6):966 | 10 AI models vs 10 dietitians | 15 hospital meals | Direct weighing | Top models often within 10% for energy, r above 0.8 | All AI models overestimated fat by more than 20% |
| Rodríguez-Jiménez et al., 2025 | Nutrients 17(22):3613 | ChatGPT-5 | 195 dishes | Recipes, SNAPMe, weighed home meals | 30.5% from photo alone; 13.9% with ingredient list | 24.4% with a description; 18.1% ingredients only |
| Fridolfsson et al., 2025 | Current Developments in Nutrition 9(10):107556 | GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro | 52 photos | Direct weighing plus nutrient database | 35.8% (GPT-4o and Claude), 64.2% (Gemini) | Protein error 60.7% to 109.9%. Underestimation grew with portion size |
| Belfarsi and Ekeocha, 2025 | Kennesaw State C-Day showcase, master's research | GPT Vision, Claude, Gemini | MenuMatch and CGMacros datasets | Nutritionist labels, MyFitnessPal logs | 13.83% (GPT Vision, best of 3) | Not peer reviewed. Reference was logged, not weighed |
| Cureus poster, 2026 | Cureus poster | GPT-5.2, Gemini 3 Flash, Claude Sonnet 4.6 | 20 meals | Kitchen-scale weights, USDA values | Reasonable averages, considerable meal-level errors | Small sample. Poster, not a paper |
| Cofre et al., 2025 | British Journal of Nutrition 133(9) | Systematic review of AI dietary assessment | 13 studies | Traditional assessment methods | 6 studies found correlation above 0.7 for calories | 61.5% of studies at moderate risk of bias |
| Shonkoff et al., 2023 | Annals of Medicine 55(2) | Review of automated AI image methods | 52 studies, 2010 to 2023 | Nutrient tables (51%), weighed food (27%) | Average calorie error 0.10% to 38.3% | Lower for simple foods. AI on par with humans |
| Hariri, 2025 | American University of Beirut, MS thesis | Recognition of Lebanese and regional dishes | 174 dish classes, 7,067 real images | Labelled images | 78.66% recognition accuracy | Recognition, not calories |
| Aktı, Qaraqe, Ekenel, 2022 | ICIAP 2021 workshop (arXiv) | Middle Eastern food recognition | 23 dish classes | Labelled images | 94% recognition accuracy | Recognition, not calories |
| SnapCalorie, 2023 | Company claims (TechCrunch, website) | SnapCalorie app | 5,000-meal training set | Not published | "Under 20%" error; "2x more accurate than nutritionists" | Vendor claims, no published test |
| Our own test, May 2026 | Callie (not independent) | Callie, Cal AI | 20 meals | 0.1 g kitchen scale, USDA FoodData Central | Callie 13%, Cal AI 17% | We make Callie. See [method](/compare/callie-vs-cal-ai) |

## Study by study

Each section gives one study's method, key number and main caveat.

### NIH/NIDDK app test (NUTRITION 2026)

**Four popular photo apps underestimated weighed meals by 252 to 345 kcal each, about a third.** It is the only test of commercial apps against a metabolic kitchen, where ingredients are weighed to 0.1 g.

NIDDK researchers ran photos of 102 meals through MyFitnessPal, Lose It!, Cal AI and Appediet. Fat came in about 30 g under per meal, and carbohydrate was the most consistent macro. [Healio](https://www.healio.com/news/primary-care/20260804/ai-photobased-calorietracking-tools-underestimate-them-by-33) reported the per-app averages: Appediet 252 kcal under, MyFitnessPal 327, Lose It! 333 and Cal AI 345. In a follow-up on 200+ more meals, keto meals did worse.

The photos were logged without corrections. Aaron Hengist, an NIDDK postdoctoral fellow, said in the [press release](https://www.eurekalert.org/news-releases/1136415): "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." Leigh A. Frame, PhD, of George Washington University, wrote in Healio's perspective that "these tools are probably best viewed as behavior-change aids rather than precision nutrition instruments."

This is a conference abstract, not yet peer reviewed. Our [write-up of the NIH study](/blog/ai-calorie-app-accuracy-study-2026) covers the method.

### Li, 802 meals in a feeding trial (2026 preprint)

**On 802 weighed meals, three of six models were accurate on average, but none was reliable meal by meal.** It is the largest sample here, a preprint posted October 19, 2026.

The [preprint](https://arxiv.org/abs/2610.26538) tested Claude and GPT models against weighed intake in a controlled feeding trial. Over 148 feeding days, three models kept the 90% confidence interval of their average error within a 10% margin. No model met the pre-set limit of 30% for individual meals. Every model underestimated protein, most in protein-dense meals.

The author's conclusion applies to every study in this table: a low average error is not evidence that single-meal estimates are accurate.

### Isobe et al., hospital meals vs dietitians (2026)

**For energy, the best AI models matched registered dietitians on 15 weighed hospital meals.** Fat was the exception: every AI model overestimated it.

This [Nutrients study](https://doi.org/10.3390/nu18060966) photographed 15 standardized hospital meals under fixed lighting. Ten registered dietitians and 10 AI models, including GPT-4o and Gemini 1.5 Pro, estimated each one. For energy and carbohydrate, the top models reached correlations above 0.8 and often landed within 10% of the weighed values.

Protein and fat were worse: all AI models overestimated lipids, with a mean bias above 20%. That runs opposite to the NIH finding: the direction of a fat error depends on the food. The authors call for "human-in-the-loop" checking and for tools to take cooking details into account.

### Rodríguez-Jiménez et al., ChatGPT-5 with more context (2025)

**ChatGPT-5 missed by 30.5% from a photo alone and by 13.9% when it also had the ingredient list.** That drop is the clearest evidence that correcting the AI's guess pays off.

The [Nutrients paper](https://pmc.ncbi.nlm.nih.gov/articles/PMC12655113/) ran 195 dishes through ChatGPT-5 four ways. Energy error was 30.5% (about 123 kcal per dish) from the photo alone, 24.4% with a short description, 13.9% (about 53 kcal) with the photo plus ingredients and amounts, and 18.1% with the ingredients but no photo. The photo helped; typed detail helped more.

This is how a review screen works. Callie shows each photo estimate and saves nothing until you confirm, so you can add the oil or fix a portion first. The study did not test Callie. It tested correction in general, and correction moved the number.

### Fridolfsson et al., GPT-4o, Claude and Gemini (2025)

**GPT-4o and Claude 3.5 Sonnet both missed calories by 35.8% on 52 food photos, and Gemini 1.5 Pro by 64.2%.** All three underestimated more as portions got bigger.

The [study in Current Developments in Nutrition](https://pmc.ncbi.nlm.nih.gov/articles/PMC12513282/) compared each model with weighed reference values for single foods and full meals in three portion sizes. Protein errors ran from 60.7% to 109.9%.

The authors found GPT-4o and Claude comparable with traditional self-reported dietary assessment, without the effort for the user. Their verdict was still cautious: "these general-purpose LLMs are not yet suitable for precise dietary assessment in clinical or athletic populations where accurate quantification is critical." See [can ChatGPT count calories](/blog/can-chatgpt-count-calories) for the chatbot route in practice.

### Kennesaw State master's research (2025)

**GPT Vision had the lowest calorie error of three models, 13.83%.** It is a student showcase project, not peer reviewed.

El Arbi Belfarsi and Henry Ekeocha [benchmarked GPT Vision, Claude and Gemini](https://digitalcommons.kennesaw.edu/cday/Spring_2025/Masters_Research/10) on two public datasets: MenuMatch, labelled by a nutritionist, and CGMacros, built from users' MyFitnessPal entries. Gemini had the most balanced macro errors: protein 12.55%, carbohydrate 19.57% and fat 17.07%.

Read the 13.83% with care: the reference was logged data, not weighed food, so the model was graded against another estimate.

### Cureus poster, three current models (2026)

**On 20 weighed meals, GPT-5.2, Gemini 3 Flash and Claude Sonnet 4.6 gave reasonable average estimates but considerable errors on single meals.** It is a poster, and the sample is small.

The [poster](https://www.cureus.com/posters/3217-validity-and-reliability-of-photo-based-ai-estimates-of-meal-calories-and-macronutrients) weighed meals on a validated digital kitchen scale and photographed each at 45 and 90 degrees with a US quarter for scale. The summary reports reasonable average nutrient estimates and relatively high internal consistency, but considerable meal-level errors. We could not confirm a single headline error figure, so we don't quote one.

### Two systematic reviews (2023 and 2025)

**A 2023 review of 52 studies found average calorie errors from 0.10% to 38.3%, with AI comparable to or better than human estimators.** Errors were lower for simple, single foods.

[Shonkoff et al.](https://pubmed.ncbi.nlm.nih.gov/38060823/) screened 14,059 papers in *Annals of Medicine* and kept 52 from 2010 to 2023. Ground truth came mostly from nutrient tables (51%) or weighed food (27%). The studies were too varied to pool, but the authors concluded that AI methods match the accuracy of human estimators and may exceed it.

A 2025 review in the [British Journal of Nutrition](https://pmc.ncbi.nlm.nih.gov/articles/PMC12229984/) by Cofre and colleagues found 13 eligible studies. Six reported correlations above 0.7 between AI and traditional methods for calories, and 61.5% had a moderate risk of bias. Both reviews mostly predate today's chatbots.

### Middle Eastern food recognition (2022 and 2025)

**Research on Middle Eastern food is thin: a 2025 thesis reached 78.66% recognition accuracy across 174 Lebanese and regional dishes.** Neither study measures calories, but an app has to recognise a dish before it can count it.

A 2025 [American University of Beirut thesis](https://scholarworks.aub.edu.lb/items/b4a5f886-4834-48a3-bdf1-31aeedde9ce3/full) by Sumayah Hariri built a 174-class set of Lebanese and regional dishes from 7,067 real images and reached 78.66% accuracy. An [earlier paper](https://www.arxiv.org/abs/2204.09432) by Aktı, Qaraqe and Ekenel reached 94% on 23 Middle Eastern classes and noted that this cuisine "has remained unexplored" in food recognition.

Products show the gap too. In 2023, SnapCalorie's co-founder told TechCrunch its algorithm may be biased toward American food. If you cook Arab or South Asian food, expect more correction work; see [tracking Indian food](/blog/calorie-tracker-for-indian-food).

### SnapCalorie's claims (vendor)

**SnapCalorie says its error is under 20% and that it is twice as accurate as nutritionists.** Both are company claims with no published test behind them.

Co-founder Wade Norris told [TechCrunch](https://techcrunch.com/2023/06/26/snapcalorie-computer-vision-health-app-raises-3m/) in June 2023: "On average, the team is able to reduce the caloric error to under 20%." TechCrunch reported the model was trained on 5,000 meals photographed on a robotic rig, with every ingredient weighed. The [SnapCalorie homepage](https://www.snapcalorie.com/) says the app is "2x more accurate than nutritionists" and links to Google's 2021 Nutrition5k research, without saying how the 2x was measured.

The numbers may be right, but until an outside team tests the app, treat them like any vendor claim, ours included. See [Callie vs SnapCalorie](/compare/callie-vs-snapcalorie).

### Our own 20-meal test (May 2026)

**In our own test, Callie averaged 13% error and Cal AI 17%.** We make Callie, so this is not independent, and it belongs at the bottom of the table.

In May 2026 we weighed 20 meals on a 0.1 g kitchen scale, photographed each one with a fork in frame, logged it once in each app and compared the result with USDA FoodData Central values. Layered dishes like lasagna and biryani were hardest for both apps, at 22% for Callie and 26% for Cal AI. The [method and results](/compare/callie-vs-cal-ai) are on our comparison page.

Two cautions. Twenty meals is a small sample. And Cal AI has since connected to MyFitnessPal's food database, which may have narrowed the gap; we have not re-run the full test.

## What the studies agree on

**The studies agree that AI estimates lean low, and that fat, big portions and mixed dishes cause most of the miss.** The pattern repeats across the table.

- **Underestimation.** All four apps in the NIH test came in low, as did all three chatbots in Fridolfsson's study. Random errors would cancel out over a week. These don't.
- **Fats and oils.** Fat was about 30 g under per meal in the NIH test. Oil soaked into food leaves no visual trace. The 2026 hospital-meal study is the exception, with fat overestimated by more than 20%.
- **Large portions.** In the 2025 chatbot study, underestimation grew with portion size (bias slopes of -0.23 to -0.50).
- **Mixed dishes.** The 2023 review found lower errors for simple, single foods. Layered dishes were the hardest category in our own test.
- **Keto meals.** The NIH follow-up on 200+ meals found the apps did worse on high-fat ketogenic meals, which fits the fat pattern.

## What the studies don't tell you yet

**The evidence is thinner than headlines suggest: small samples, unreviewed abstracts and models that change every few months.** Treat each number as a snapshot.

- **Small samples.** Most studies used 15 to 195 meals. Only the 2026 preprint passed 800.
- **Abstracts and preprints.** The NIH test, the 802-meal study, the Kennesaw project and the Cureus poster have not been through journal peer review.
- **Moving targets.** The 2025 papers tested GPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro, all since replaced. Apps update models without notice, so a 2025 figure may not describe your app today.
- **Averages hide misses.** As the 802-meal preprint showed, a model can be close on average and still miss single meals widely.
- **Real use.** No study yet measures what people log after editing an estimate in a commercial app.

## How to get closer to the real number

**The research points to one habit above all: check and correct the estimate before you save it.** Typed detail cut ChatGPT-5's error from 30.5% to 13.9%.

1. **Add the fat yourself.** Log cooking oil, butter and dressing as their own items. One tablespoon of oil is about 120 kcal, and no camera sees it. Our guide to [hidden calories in oil, butter and sauces](/blog/hidden-calories-oil-butter-sauces) lists the worst offenders, and [why AI calorie apps undercount](/blog/why-ai-calorie-apps-underestimate) walks through each cause with a fix.
2. **Fix the portion, keep the food.** Recognition is usually right. Quantity is where estimates drift, especially on big plates.
3. **Describe mixed dishes.** For a curry or casserole, add the main ingredients and amounts, as in the 13.9% condition.
4. **Weigh only the dense foods.** A scale helps most with oil, nut butter, cheese, granola and pasta. See [do you need a food scale with an AI tracker](/blog/do-you-need-a-food-scale-with-ai-tracker).
5. **Check your weight trend.** If your log shows a deficit and the scale hasn't moved in three weeks, the log is probably low.

## When a non-AI tracker is the better choice

**If you weigh everything anyway, a database tracker such as Cronometer is likely to beat any photo estimate.** The studies above measure guesses, and a weighed entry isn't a guess.

Cronometer's free tier logs by database search and barcode scan, tracks a long list of nutrients and keeps AI photo logging in its paid Gold plan. That suits people on a medical diet, athletes making weight and anyone who already likes using a scale. Our list of the [best calorie trackers without AI](/blog/calorie-tracker-without-ai) covers Cronometer, MyNetDiary and two open-source apps.

AI photo logging makes more sense when speed decides whether you log at all. A corrected estimate 14% off beats a meal never recorded.

This article is general information, not medical advice. If you have a medical condition, take medication affected by diet, are pregnant, or have a history of disordered eating, talk to a qualified healthcare professional before changing how you eat or track food.

## Sources

1. NIDDK / American Society for Nutrition press release via EurekAlert! (July 25, 2026). "Photo-based calorie-tracking apps may underestimate energy in meals." https://www.eurekalert.org/news-releases/1136415
2. Healio (August 4, 2026). "AI photo-based calorie-tracking tools underestimate them by 33%," with a perspective from Leigh A. Frame, PhD. https://www.healio.com/news/primary-care/20260804/ai-photobased-calorietracking-tools-underestimate-them-by-33
3. Nutritional Outlook (2026). "Study finds popular AI calorie tracking apps underestimate meal energy by about a third." https://www.nutritionaloutlook.com/view/study-finds-popular-ai-calorie-tracking-apps-underestimate-meal-energy-by-about-a-third
4. HealthDay (July 27, 2026). "How Accurate Are Photo-Based Calorie Apps? 4 Are Put To The Test." https://www.healthday.com/health-news/nutrition/how-accurate-are-photo-based-calorie-apps-4-are-put-to-the-test
5. Li X (October 19, 2026). "Small Average Error, Wide Individual Error: General-Purpose LLMs Estimating Energy Intake from Meal Photos in a Feeding Trial." arXiv preprint 2610.26538. https://arxiv.org/abs/2610.26538
6. Isobe T, Zhang LW, Murakami H, et al. (2026). "Accuracy of AI-Based Nutrient Estimation from Standardized Hospital Meal Images: A Comparison with Registered Dietitians." *Nutrients* 18(6):966. https://doi.org/10.3390/nu18060966
7. Rodríguez-Jiménez M, Martín-del-Campo-Becerra GD, Sumalla-Cano S, Crespo-Álvarez J, Elio I (November 19, 2025). "Image-Based Dietary Energy and Macronutrients Estimation with ChatGPT-5: Cross-Source Evaluation Across Escalating Context Scenarios." *Nutrients* 17(22):3613. https://pmc.ncbi.nlm.nih.gov/articles/PMC12655113/ and https://www.mdpi.com/2072-6643/17/22/3613
8. Fridolfsson J, Sjöberg E, Thiwång M, Pettersson S (2025). "Performance Evaluation of 3 Large Language Models for Nutritional Content Estimation from Food Images." *Current Developments in Nutrition* 9(10):107556. https://pmc.ncbi.nlm.nih.gov/articles/PMC12513282/
9. Belfarsi EA, Ekeocha H (April 15, 2025). "Assessing the Performance of Intelligent Agents in Visual Food Recognition Relative to Manual Data Entry." Kennesaw State University C-Day Computing Showcase. https://digitalcommons.kennesaw.edu/cday/Spring_2025/Masters_Research/10
10. Cureus poster. "Validity and Reliability of Photo-Based AI Estimates of Meal Calories and Macronutrients." https://www.cureus.com/posters/3217-validity-and-reliability-of-photo-based-ai-estimates-of-meal-calories-and-macronutrients
11. Cofre S, Sanchez C, Quezada-Figueroa G, López-Cortés XA (2025). "Validity and accuracy of artificial intelligence-based dietary intake assessment methods: a systematic review." *British Journal of Nutrition* 133(9):1241–1253. https://pmc.ncbi.nlm.nih.gov/articles/PMC12229984/
12. Shonkoff E, Cara KC, Pei XA, et al. (2023). "AI-based digital image dietary assessment methods compared to humans and ground truth: a systematic review." *Annals of Medicine* 55(2):2273497. https://pubmed.ncbi.nlm.nih.gov/38060823/
13. Hariri S (2025). "Estimating Calories and Macronutrients from Lebanese Food Images Using Deep Learning–Based Classification and Single-View Portion Size Estimation." MS thesis, American University of Beirut. https://scholarworks.aub.edu.lb/items/b4a5f886-4834-48a3-bdf1-31aeedde9ce3/full
14. Aktı Ş, Qaraqe M, Ekenel HK (2022). "A Mobile Food Recognition System for Dietary Assessment." ICIAP 2021 workshop, arXiv 2204.09432. https://www.arxiv.org/abs/2204.09432
15. TechCrunch (June 26, 2023). "SnapCalorie taps AI to estimate the caloric content of food from photos." https://techcrunch.com/2023/06/26/snapcalorie-computer-vision-health-app-raises-3m/
16. SnapCalorie homepage (accessed October 20, 2026). https://www.snapcalorie.com/
17. Callie internal 20-meal kitchen-scale benchmark (May 2026). Methodology at https://www.mycallie.app/en/compare/callie-vs-cal-ai

## Frequently asked questions

**Q: How accurate are AI calorie apps?**
A: Typically 14% to 36% off for calories, based on studies published from 2023 to 2026. From a photo alone, error sits near the top of that range: an NIH team found four popular apps underestimated meals by about a third in 2026, and chatbots missed by 30.5% to 35.8% in 2025 studies. When a person corrects the ingredients and portions, error drops toward the low end.

**Q: Do AI calorie apps overestimate or underestimate?**
A: Mostly underestimate. All four apps in the 2026 NIH test came in low, by 252 to 345 kcal per meal, and a 2025 study found three chatbots underestimated more as portions got bigger. The usual cause is fat and oil the camera cannot see. One exception: a 2026 study of 15 hospital meals found AI models overestimated fat by more than 20%.

**Q: Which AI calorie app was most accurate in the NIH study?**
A: According to Healio's report of the abstract, average underestimates were 252 kcal for Appediet, 327 kcal for MyFitnessPal, 333 kcal for Lose It! and 345 kcal for Cal AI. All four erred in the same direction. The coverage does not say whether the gaps between apps are statistically significant, and the work has not yet been peer reviewed.

**Q: Is ChatGPT as accurate as a calorie app at estimating calories?**
A: From a photo alone, roughly the same range. ChatGPT-5 averaged 30.5% error on 195 dishes in a 2025 Nutrients study, close to the one-third underestimate the NIH team found for dedicated apps. ChatGPT got much better when given ingredient amounts, but it keeps no food diary or daily totals.

**Q: Does correcting the AI's estimate actually help?**
A: Yes. In the 2025 ChatGPT-5 study, adding an ingredient list with amounts cut energy error from 30.5% to 13.9%, or from about 123 to 53 kcal per dish. The NIH researchers also warned that people who do not adjust portions should take the results 'with a grain of salt'.

**Q: Are these AI calorie studies peer reviewed?**
A: Some are. The ChatGPT, Claude and Gemini studies in Current Developments in Nutrition and Nutrients, the two systematic reviews and the 2026 hospital-meal study are peer reviewed. The NIH app test is a conference abstract, the 802-meal study is a preprint, and the Kennesaw State and Cureus results are a student showcase and a poster.