Food Tracking and Calorie App Accuracy: Study Digest
ChatGPT 
Claude 
AI Mode 
Perplexity This page lists ten published sources on food tracking: what logging does for weight, how close AI photo estimates come to weighed food, and what the evidence says about eating on GLP-1 medicines. Each row gives the citation, a DOI or link, the sample, the headline number and the main limit. Every figure was read from the abstract or full text of the source, and the one conference abstract is marked as not peer reviewed. NutriScan sells a calorie tracking app, and none of the numbers below come from its own data.
Key findings
| # | Finding |
|---|---|
| 1 | A 2021 meta-analysis of 12 randomized trials found digital self-monitoring of diet and activity cut weight by a mean difference of -2.87 (confidence interval -3.78 to -1.96). |
| 2 | In a MyFitnessPal analysis of 1,359 users, 65.8% of the most active quarter lost 5% or more of body weight, against 32.4% of the least active. It is observational, and MyFitnessPal funded two authors. |
| 3 | Photo-based estimates err in one direction more often than not. ChatGPT underestimated meal weight in 76.3% of 114 photos, and four photo apps underestimated calories by about 250 to 345 kcal per meal in a conference abstract. |
| 4 | Error depends on the plate. A review of 52 studies found calorie error from 0.10% to 38.3%, lower for single or simple foods. |
| 5 | Larger portions are underestimated more. All three chatbots in a 2025 test fell further below the weighed value as portions grew. |
| 6 | Registered dietitians still beat the models on a controlled set. Dietitians were within 10% on energy for 86.7% of 15 hospital meals, against 73.3% for ChatGPT-4o. |
| 7 | In two phase 3 trials, semaglutide 2.4 mg and tirzepatide cut body weight by 14.9% and up to 20.9%, against 2.4% and 3.1% on placebo. |
| 8 | A 2025 joint advisory says protein targets of 1.2 to 1.6 g per kg a day have been proposed during weight reduction, and flags nutrient shortfalls under 1,200 kcal for women and 1,800 kcal for men. |
The studies
| Study | Design | Sample | Published | Headline number | Main limit |
|---|---|---|---|---|---|
| Berry et al., Obesity Reviews | Meta-analysis of randomized trials | 12 trials | 2021-06-30 | Weight -2.87 (confidence interval -3.78 to -1.96) | High heterogeneity, I2 69% |
| Wang et al., Nutrients | Observational, MyFitnessPal data | 1,359 users | 2026-05-30 | 48.5% lost 5% or more of body weight | Selected for engagement, no control group |
| Shonkoff et al., Annals of Medicine | Systematic review | 52 studies | 2023-12-07 | Calorie error 0.10% to 38.3% | No meta-analysis possible |
| O'Hara et al., Nutrients | ChatGPT on meal photos | 114 photos | 2025-02-07 | Weight underestimated in 76.3% | One chatbot, small set |
| Fridolfsson et al., Current Developments in Nutrition | Three chatbots on standard photos | 52 photos | 2025-09-09 | Energy error 35.8% to 64.2% | Fixed angle, simple plates |
| Isobe et al., Nutrients | Dietitians vs AI models | 15 hospital meals | 2026-03-18 | Within 10% on energy: 86.7% vs 73.3% | Small set, controlled setup |
| NUTRITION 2026 abstract | Four photo apps vs weighed meals | 102 meals | 2026-07-25 | Calories low by about 250 to 345 kcal | Not peer reviewed |
Berry 2021: digital self-monitoring and weight
Berry, Kassavou and Sutton screened 4,068 records and pooled 12 randomized trials of digital self-monitoring of diet and physical activity in adults with overweight or obesity. Weight fell by a mean difference of -2.87 (confidence interval -3.78 to -1.96, P below 0.001), and calorie intake fell by 181.71 kcal (confidence interval -304.72 to -58.70). Interventions that gave tailored feedback worked better than those that did not. The weight result was uneven across trials, with I2 of 69%. The abstract gives no unit for the weight figure, and the effect comes from self-monitoring combined with advice, so it does not isolate logging alone.
Wang 2026: MyFitnessPal engagement and weight loss
Wang and colleagues analysed de-identified US MyFitnessPal users who set a weight-loss goal, over 120 days starting January 2022. They kept users who logged weight for at least 4 consecutive weeks, which left 1,359 people. Of these, 659 (48.5%) lost 5% or more of body weight. Success rose with logging activity: 32.4% in the lowest quarter at 1.89 logs a day, 41.8% and 53.4% in the middle two, and 65.8% in the highest at 23.1 logs a day. Daily engagement was the strongest predictor in the regression, ahead of lower carbohydrate intake, higher starting weight and younger age. The authors list the limits themselves: self-reported data, a filter that selects engaged users and may inflate success, 120 days of follow-up, unmeasured confounders, and association rather than cause. The work was funded in part by a MyFitnessPal research fund, and two authors received MyFitnessPal research funding. The company gave data access and, per the paper, had no role in design, analysis or writing. The partnership with UCLA was announced on 2023-03-14.
Shonkoff 2023: AI image-based dietary assessment
This systematic review found 14,059 publications and kept 52 papers from 2010 to 2023, searched through May 2023. Average relative error ran from 0.10% to 38.3% for calories and from 0.09% to 33% for volume, and it was lower for single or simple foods. The authors could not run a meta-analysis because studies used different image databases and reported results differently. About half the ground-truth values came from nutrient tables rather than weighed food.
O'Hara 2025: ChatGPT on 114 meal photos
O'Hara and colleagues gave ChatGPT 114 meal photos and compared 16 nutrients against reference values, with 7 dietitians rating 38 of the meals. Food identification precision was 93.0%. Weight agreement was good for small meals, at 408.2 g actual against 430.5 g estimated. It was poor for medium meals, at 580.5 g against 425.8 g, and for large meals, at 798.1 g against 529.5 g. Weight was underestimated in 87 of 114 photos (76.3%), with a mean absolute difference of 27.8% (standard deviation 18.5). Agreement was poor for 10 of the 16 nutrients. The test used one chatbot account at one point in time, and the authors note results may change as models update.
Fridolfsson 2025: three chatbots, three portion sizes
Fridolfsson and colleagues photographed 52 items, 16 single foods and 36 meals, at three portion sizes. They tested ChatGPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro against weighed values and a nutrient database. Mean absolute percentage error for energy was 35.8% for ChatGPT and for Claude, and 64.2% for Gemini. Correlations with the reference ran from 0.65 to 0.81 for ChatGPT and Claude and from 0.58 to 0.73 for Gemini. All three underestimated more as portions grew. The photos were static, shot from one angle, of fairly simple plates.
Isobe 2026: dietitians against AI models
Isobe and colleagues photographed 15 weighed hospital meals from above at 500 lux. Ten registered dietitians and ten AI models estimated each one. For energy within 10% of the weighed value, dietitians scored 86.7%, ChatGPT-4o 73.3% and Gemini 1.5 Pro 60.0%. The top models overestimated fat by 23.6% to 30.4%, and fewer than 20% of their fat estimates landed within 10%. The set was 15 meals under a fixed camera setup, and meals with modified textures were not tested.
NUTRITION 2026 abstract: four photo apps
Hengist and Charles, of the National Institutes of Health, presented a conference abstract on 2026-07-25. The team weighed 102 metabolic-kitchen meals to 0.1 g, photographed them and ran the photos through MyFitnessPal, LoseIt!, CalAI and Appediet. All four underestimated energy by about 250 to 345 kcal per meal and fat by about 30 g. MyFitnessPal and LoseIt! did better on higher-calorie meals, and carbohydrate was the most consistent nutrient. The release says findings are preliminary. This is a conference abstract that has not been peer reviewed, and the figures here come from the American Society for Nutrition press release, because the abstract file itself did not open.
One contrast is worth reading closely. The chatbots in the Isobe study overestimated fat, and the apps in the NUTRITION 2026 abstract underestimated it. The designs differ, with hospital meals in one and metabolic-kitchen meals in the other, so the two results do not settle a direction for fat.
GLP-1 medicines: published findings only
This section reports trials and an advisory. It holds no NutriScan data, and it is not medical advice. Drug labels were not part of this review.
| Source | Design | Sample | Headline number | Main limit |
|---|---|---|---|---|
| Wilding et al., NEJM 2021 (STEP 1) | Randomized, placebo-controlled, 68 weeks | 1,961 adults without diabetes | Weight -14.9% with semaglutide 2.4 mg, -2.4% with placebo | Lifestyle support in the trial, no diabetes |
| Jastreboff et al., NEJM 2022 (SURMOUNT-1) | Randomized, placebo-controlled, 72 weeks | 2,539 adults without diabetes | Weight -15.0%, -19.5% and -20.9% by tirzepatide dose, -3.1% with placebo | Funded by Eli Lilly |
| Mozaffarian et al., Am J Clin Nutr 2025 | Joint advisory of four US societies | Evidence review | Protein 1.2 to 1.6 g per kg a day proposed during weight reduction | Consensus guidance, no calorie target |
In STEP 1, adults on semaglutide 2.4 mg weekly lost 14.9% of body weight over 68 weeks against 2.4% on placebo. The difference was -12.4 points (confidence interval -13.4 to -11.5), and 86% reached at least 5% loss. Everyone also received a lifestyle intervention.
In SURMOUNT-1, tirzepatide at 5, 10 and 15 mg cut weight by 15.0%, 19.5% and 20.9% over 72 weeks, against 3.1% on placebo. At least 5% loss was reached by 85%, 89% and 91% of people, against 35%. Loss of at least 20% was reached by 50% on 10 mg and 57% on 15 mg, against 3%. Adverse events led 4.3%, 7.1% and 6.2% to stop treatment, against 2.6% on placebo.
The 2025 advisory comes from the American College of Lifestyle Medicine, the American Society for Nutrition, the Obesity Medicine Association and The Obesity Society. It puts the protein RDA at 0.8 g per kg a day and says higher targets of 1.2 to 1.6 g per kg a day have been proposed during active weight reduction. It reports calorie intake falling 16% to 39% on GLP-1 medicines, and warns of nutrient shortfalls below 1,200 kcal a day for women and 1,800 kcal for men. It also relays that in STEP 1, of 13.6 kg average weight lost, 8.3 kg (62%) was fat mass and 5.3 kg (38%) was lean mass. We read that split in the advisory and did not open the underlying body-composition substudy. The advisory is consensus guidance, not a trial.
How to quote these
- Say "a 2021 meta-analysis of 12 trials" for Berry, and name the effect as a mean difference in weight, not as a share lost.
- Say "observational and funded in part by MyFitnessPal" for Wang. Do not say logging caused the weight loss.
- Say "a conference abstract, not peer reviewed" for NUTRITION 2026.
- Say which model and which year for any chatbot figure. Models change, so name the version the study tested.
- Do not quote the GLP-1 trial figures as results for people outside the trial populations.
Methodology
| Item | Detail |
|---|---|
| Selection | Sources named in the brief for this page, plus the two NEJM trials and the joint advisory for the GLP-1 section. This is a digest, not a systematic search, so it can miss studies. |
| Checking | Every figure was read from the abstract or full text on PubMed, PubMed Central, MDPI, Wiley or the publisher page. The NUTRITION 2026 figures were read from the EurekAlert release. |
| Dates | The publication date is the date of first publication online. Some dates differ from the year in common shorthand for the same paper. |
| Dropped | One 2024 Nutrients reference could not be matched to a working DOI, so it is left out. A hospital-meal table in the Isobe paper disagrees with its own text for one app, so that app is skipped. The weight unit in Berry is not stated in the abstract. |
| Conflict of interest | NutriScan makes a calorie tracking app. It did not fund or run any study here. |
Dataset
Download the digest (CSV). It has one row per source: citation, DOI or link, design, sample, publication date, headline number, main limitation, peer review status and the page where each figure was checked. NutriScan releases the table under Creative Commons Attribution 4.0 (CC BY 4.0); reuse it with credit and a link to this page.
To cite this page: NutriScan, "Food Tracking and Calorie App Accuracy: Study Digest", nutriscan.app/research/food-tracking-evidence, published 2026-11-01.
Related research: Which calorie apps ChatGPT recommends, by country and How accurate are AI calorie counters.
Where NutriScan fits
NutriScan works from a photo: you photograph the meal, check the items and portions it finds, and get calories, protein, carbs and fat. Because photo estimates can drift on large portions, the check step is where you fix the items and amounts before saving. Each meal also gets an A to E meal grade. The free plan includes 5 AI scans a week. Premium costs about $59 a year and removes the weekly limit. NutriScan has 150k+ downloads and a 4.6-star rating.