Skip to content
LifeSugar

How accurate is AI at counting carbs from a photo?

8 min readLifeSugar team

You take a photo of your plate and the app tells you how many carbs it has. Sounds good. The question is whether you can trust it. At LifeSugar we use Gemini, from Google, for exactly that, and one of us counts carb portions at every meal. So we've read the studies published up to October 2026. We tell you what they say, including what doesn't make us look good, and how they have shaped the way we designed the app.

In short

  • From a photo alone, AI is usually off by 10 to 28 g of carbs per meal 1, 9, 13. That is, between one and three portions.
  • People get it wrong too: around 15 g per meal on average in adults with type 1 15.
  • What helps it most is telling it how much there is. With the ingredients and their grams, or a detailed description, the error drops to 2–6 g 2, 9. With the real weight, it also drops a lot 10.
  • It does well with an apple. It struggles more with a stew or a large plate 3, 7.
  • No study says it is safe to calculate the bolus from what a photo estimates.

How far off is AI when counting carbs from a photo?

From the photo alone, between 10 and 28 g of carbs per meal on average, depending on the study and the model: between one and three portions 1, 9, 13. The best result among the studies we cite, 6.6 g, is from Gemini-3 Pro in a small study 2.

These are the studies that compare AI with a flesh-and-blood person, or with the real weight of the food. “Mean error” is how far the estimate is from the real carbs, on average, at each meal.

With the photo alone, AI is usually off by more than one portion. If you give it the amount, less

Mean error per meal, in grams of carbs. The shorter the line, the better.

  • People
  • AI
  • Within one portion

Goncalves, 2026 130 photographed meals, against 6 dietitians

  • Dietitians
  • ChatGPT
  • Claude
  • Gemini

Soyer, 2026 218 meals and 4 snacks, against 14 highly trained adults with type 1

  • Adults with type 1
  • Gemini-3 Pro, photo only
  • Gemini-3 Pro, photo and description

Rodríguez-Jiménez, 2025 9195 dishes, giving the AI more and more context

  • ChatGPT-5, photo only
  • ChatGPT-5, with ingredients and amounts
Each block is a different study, with its own meals, method and model version: compare within each block, not between blocks.
View the data in a table
Mean error per meal, in grams of carbs. The shorter the line, the better.
StudyWhoMean errorIn portions
Goncalves, 2026Dietitians13 g1.3
Goncalves, 2026ChatGPT20 g2
Goncalves, 2026Claude23 g2.3
Goncalves, 2026Gemini28 g2.8
Soyer, 2026Adults with type 18.3 g0.8
Soyer, 2026Gemini-3 Pro, photo only6.6 g0.7
Soyer, 2026Gemini-3 Pro, photo and description2.1 g0.2
Rodríguez-Jiménez, 2025ChatGPT-5, photo only11.7 g1.2
Rodríguez-Jiménez, 2025ChatGPT-5, with ingredients and amounts6.2 g0.6

There is one finding in the first study that matters a lot to us. Gemini overshot by 20 g or more in 38% of meals. The dietitians, in 3% 1. Overestimating is what leads to taking too much insulin.

Against apps built specifically for carb counting, the result is similar: in 246 hospital meals, the apps were off by 13 to 21 g and ChatGPT-5 by 18 g 4. In that study it isn't clear whether ChatGPT was given the photo or a description.

The second study is the one that treats AI best, but its authors urge caution: it's 14 people, highly trained 2. And all of this is very recent. The latest systematic review on estimating food from images is from 2023 and doesn't include a single study with these models 14.

How much error is acceptable when counting carbs?

About 10 g, one portion. In children and teenagers on intensive insulin, being off by 10 g didn't change post-meal blood glucose; being off by 20 g did 18, 19.

Ten grams of carbs is one portion. It's the measure we use throughout the article.

In children and teenagers on intensive insulin, being off by 10 g didn't change post-meal blood glucose 18. Being off by 20 g did: when 20 g too many were counted, one in three meals ended in hypoglycaemia 19.

So we're not looking for an AI that gets it right to the gram. We're looking to find out how often it stays within that one portion of margin.

Does AI count carbs better than a person with diabetes?

It depends on the study, and the evidence is mixed. People with diabetes get it wrong too: around 15 g per meal on average in adults with type 1 15.

If you count carbs, you already know: it's hard.

In 50 adults with type 1, the mean error was 15.4 g per meal. And whoever was more off had more variable blood glucose 15. In another study, with 61 adults on insulin, the mean score was 59% and 82% overcounted 16. In children and their families, 73% of estimates fell within 10–15 g 20.

Against dietitians, AI was off more in the French study 1. Against 14 highly trained patients, Gemini-3 Pro was off slightly less 2. So you can't say AI counts better than you. You can say nobody starts from zero errors.

On which dishes is AI most off?

On dishes with many ingredients, on large portions and when the photo changes angle 3, 5, 7.

The studies agree on three trouble spots:

The blue line on the rice is what the AI calculates; the heap is what is there. The blue cone is what the phone sees.

  1. Dishes with lots of things. With fruit and vegetables, ChatGPT-4o stayed within 10 g in 93% of cases. With composite dishes, in 47% 3. Rice, pasta, potato and mash are the hardest, even for dietitians 17.
  2. Large portions. The fuller the plate, the more it falls short 7, 8. It recognises well what is there (it got 93% of foods right in an Irish study), but not how much 8.
  3. The same photo, a different figure. If you change the angle, or repeat the same photo, the number changes. And a newer model isn't more stable: GPT-5.1 varied more than GPT-4o 5. The models don't agree with each other either 6.

How can the accuracy of AI when counting carbs be improved?

By giving it the amount. With the ingredients and their grams, or a detailed description of the dish, the error drops to 2–6 g per meal 2, 9. With the real weight, it also drops a lot 10.

Knowing what is on the plate is not the problem. The problem is how much there is.

When the AI is given the real weight, Gemini 2.5 Flash's error falls from 57% to 20% 10. With the list of ingredients and their amounts, ChatGPT-5 goes from 11.7 g to 6.2 g 9. Telling it the name of the dish or where you're eating it also helps a little 12, 13.

What hasn't been shown is that putting a card or a fork next to the plate to give a sense of scale helps. In composite dishes it didn't improve anything 3.

Which AI counts carbs best: ChatGPT, Gemini or Claude?

It depends on the version, not the brand. Gemini appears among those that are most off in some studies and among those that are least off in others 1, 2, 13.

LifeSugar uses Gemini, so here we have to be clear.

In a Swedish study, Gemini 1.5 Pro deviated by 66% on average on carbs; ChatGPT-4o, by 48% 7. Gemini, with no version stated, was the one most off in the French study 1. On Japanese hospital menus, Gemini 1.5 Pro was within 10% in 47% of cases; the dietitians, in 73% 11.

By contrast, Gemini-3 Pro did best in the Turkish study 2, and Gemini 2.5 Pro was the best on carbs in a test with 806 photos 13.

Nobody has studied LifeSugar specifically yet. None of these figures is our app's.

Can you calculate an insulin bolus from a photo?

No study supports it. All those carried out with people with diabetes see AI as support, not as a replacement, and we haven't found any that tries dosing insulin with what a photo estimates 1, 2, 3, 4.

That's why LifeSugar doesn't calculate boluses. It gives you the carbs and portions, and you decide the dose with your healthcare team. LifeSugar is not a medical device.

Which studies is LifeSugar based on?

The ones in this article. Every decision in the app answers a published result: correcting before saving, being able to say the amount, not calculating the bolus and seeing the meal next to blood glucose.

Reading all this led us to design the app like this:

  • Everything can be corrected before saving: the weight, the carbs and the other values.

    Why
    The same meal gives different figures depending on the photo, and large portions fall short.
    Source
    5, 7, 8

  • You can tell Glucy how much you served yourself and it redoes the sum.

    Why
    Giving the amount is what most reduces the AI's error.
    Source
    2, 9, 10

  • It gives you grams and portions, not a dose.

    Why
    The margin that doesn't alter blood glucose is about 10 g, one portion, and no study supports dosing from a photo.
    Source
    18, 19

  • It saves each meal next to the blood glucose from before and after, and you can share it with your doctor.

    Why
    Being more off when counting is associated with more variable blood glucose: seeing it together helps you review it with your healthcare team.
    Source
    15

Our conclusion: a photo doesn't replace your judgement. It gives you a starting point without getting the scales out, and the last word is still yours.

How LifeSugar counts carbsWho is behind LifeSugar

References

  1. 1.Chat, Gemini and Claude at the dinner table: assessing general-purpose AI tools for carbohydrate counting in the context of type 1 diabetes. Goncalves S, Coelho C, Pretre L, et al. Diabetes Res Clin Pract. 2026;231:113031. doi.org/10.1016/j.diabres.2025.113031
  2. 2.Evaluation of large language models for carbohydrate counting in type 1 diabetes mellitus: a comparative study with patient estimations. Soyer AK, Akarsu GÇ, Yelli BD, et al. Diabetes Res Clin Pract. 2026;236:113279. doi.org/10.1016/j.diabres.2026.113279
  3. 3.Evaluating Accuracy of ChatGPT-4o in Automated Carbohydrate Estimation From Images as a Self-Management Tool for Adolescents With Type 1 Diabetes. Johansen AR, Laursen IK, Jacobsen V, et al. J Diabetes Sci Technol. 2026. doi.org/10.1177/19322968261424270
  4. 4.Comparative accuracy of smartphone apps and a generative AI tool for carbohydrate counting: An independent bicentric study. Joubert M, Dreves B, Arnould T, et al. Diabetes Obes Metab. 2026;28(3):2105-2112. doi.org/10.1111/dom.70396
  5. 5.Toward Robust AI-Assisted Dietary Assessment for Diabetes Self-Management: Quantifying and Decomposing Large Language Model Prediction Variability From Meal Images. Wang Z, Lane D, Waki K. JMIR Diabetes. 2026;11:e102715. doi.org/10.2196/102715
  6. 6.Carbohydrate counting in traditional Turkish fast foods for individuals with type 1 diabetes: Can artificial intelligence models replace dietitians?. Özkaya V, Eren E, Özgen Özkaya Ş, Özkaya G. Nutrition. 2026;142:112986. doi.org/10.1016/j.nut.2025.112986
  7. 7.Performance Evaluation of 3 Large Language Models for Nutritional Content Estimation from Food Images. Fridolfsson J, Sjöberg E, Thiwång M, Pettersson S. Curr Dev Nutr. 2025;9(10):107556. doi.org/10.1016/j.cdnut.2025.107556
  8. 8.An Evaluation of ChatGPT for Nutrient Content Estimation from Meal Photographs. O'Hara C, Kent G, Flynn AC, Gibney ER, Timon CM. Nutrients. 2025;17(4):607. doi.org/10.3390/nu17040607
  9. 9.Image-Based Dietary Energy and Macronutrients Estimation with ChatGPT-5: Cross-Source Evaluation Across Escalating Context Scenarios. Rodríguez-Jiménez M, Martín-Del-Campo-Becerra GD, Sumalla-Cano S, et al. Nutrients. 2025;17(22):3613. doi.org/10.3390/nu17223613
  10. 10.Benchmarking and Improving Foundation Model Dietary Estimates from Meal Images. Mu Y, Sun J, He J. ACM-BCB 2025. doi.org/10.1145/3765612.3767255
  11. 11.Accuracy of AI-Based Nutrient Estimation from Standardized Hospital Meal Images: A Comparison with Registered Dietitians. Isobe T, Zhang LW, Murakami H, et al. Nutrients. 2026;18(6):966. doi.org/10.3390/nu18060966
  12. 12.Customized multimodal Diabot-GPT-4o enhances accuracy of image-based dietary assessments in dietetic trainees in Taiwan: validation against weighed food records. Chen YJ, Chang CC, Hoang YN, et al. Am J Clin Nutr. 2025;122(6):1836-1849. doi.org/10.1016/j.ajcnut.2025.10.013
  13. 13.Comprehensive Evaluation of Large Multimodal Models for Nutrition Analysis: A New Benchmark Enriched with Contextual Metadata. Coburn B, He J, Rollo ME, et al. arXiv:2507.07048, 2025 (preprint). arxiv.org/abs/2507.07048
  14. 14.AI-based digital image dietary assessment methods compared to humans and ground truth: a systematic review. Shonkoff E, Cara KC, Pei XA, et al. Ann Med. 2023;55(2):2273497. doi.org/10.1080/07853890.2023.2273497
  15. 15.Carbohydrate counting accuracy and blood glucose variability in adults with type 1 diabetes. Brazeau AS, Mircescu H, Desjardins K, et al. Diabetes Res Clin Pract. 2013;99(1):19-23. doi.org/10.1016/j.diabres.2012.10.024
  16. 16.Accuracy of Carbohydrate Counting in Adults. Meade LT, Rushton WE. Clin Diabetes. 2016;34(3):142-147. doi.org/10.2337/diaclin.34.3.142
  17. 17.A Comparative Study on Carbohydrate Estimation: GoCARB vs. Dietitians. Vasiloglou MF, Mougiakakou S, Aubry E, et al. Nutrients. 2018;10(6):741. doi.org/10.3390/nu10060741
  18. 18.Children and adolescents on intensive insulin therapy maintain postprandial glycaemic control without precise carbohydrate counting. Smart CE, Ross K, Edge JA, et al. Diabet Med. 2009;26(3):279-285. doi.org/10.1111/j.1464-5491.2009.02669.x
  19. 19.In children using intensive insulin therapy, a 20-g variation in carbohydrate amount significantly impacts on postprandial glycaemia. Smart CE, King BR, McElduff P, Collins CE. Diabet Med. 2012;29(7):e21-4. doi.org/10.1111/j.1464-5491.2012.03595.x
  20. 20.Can children with Type 1 diabetes and their caregivers estimate the carbohydrate content of meals and snacks?. Smart CE, Ross K, Edge JA, et al. Diabet Med. 2010;27(3):348-353. doi.org/10.1111/j.1464-5491.2010.02945.x