Your own guess, a nutrition label, a barcode scan and a photo app can give four different carb counts for one meal. Here is how far off each tends to be, and how much the gap matters if you dose insulin for carbs.
A bowl of chili over rice can carry three different carb counts at once. There’s your own guess, the figure you get from a label or a barcode scan, and the estimate a photo app gives you. In a Montreal study, adults with long experience of type 1 diabetes misjudged their own meals by about 15 g on average, roughly a fifth of each meal 1.
If you match mealtime insulin to carbohydrate, those grams feed straight into a dose. Controlled studies in children found that a 10 g miss on one meal changed little, while a 20 g miss raised the chance of a low or a high 2, 3. Below is what research shows about each counting method, who was studied and how strong the evidence is. If you are new to counting, start with how to count carbs.
An insulin-to-carbohydrate ratio turns a carb count into a dose, so any counting error travels into the dose. With a ratio of 1 unit for every 10 g, for example, a 20 g miss shifts the calculated dose by 2 units. Fat, protein, activity and insulin already on board matter too.
Researchers in Newcastle, Australia, tested how much counting error a single dose can absorb. In their first crossover study, 31 children and teens aged 9.5 to 16.8 years ate lunches containing 50, 60 or 70 g of carbohydrate on three days in random order, each time taking the dose their usual ratio gave for 60 g 2. The 10 g swings made no significant difference to glucose levels over the next 2.5 hours or to the number of lows. After that, the 70 g lunch ran higher, by up to 1.9 mmol/L (34 mg/dL) at 3 hours, though levels stayed within international targets for after meals 2.
Their follow-up study widened the range 3. Thirty-four young people aged 8.5 to 17.7 years took a dose for 60 g and then ate breakfasts with 40, 50, 60, 70 or 80 g of carbohydrate, matched for fat, protein and fibre. The 40 g breakfast caused significantly more hypoglycemia, with about a one-in-three chance of a low between 2 and 3 hours after eating. The 80 g breakfast left glucose significantly higher than the 60 g and 70 g breakfasts from 150 to 180 minutes, and readings of 12 mmol/L (216 mg/dL) or more were more likely 3. The authors concluded that estimates should fall within 10 g of the true amount 3. Both were single-meal tests in young people, so the margin may differ for adults or larger meals.
At a hospital in Abu Dhabi, 30 children and teens aged 8 to 18 on insulin pumps logged their meals, carb estimates and glucose 2 hours after eating for 3 to 5 days, and a dietitian recounted every meal 4. Estimates within 20% of the dietitian’s figure counted as accurate, which was true for 165 of 247 meals (67%). Of the 63 underestimated meals, 55 were followed by glucose above target, and 12 of the 19 overestimated meals were followed by a low 4. It was an observational study, so other factors may have contributed.
In adults, the Montreal study found that people whose estimates were further off had more variable glucose and spent less time between 4.0 and 10.0 mmol/L (72-180 mg/dL) 1. The link was modest, with counting error explaining about 11% of the differences in those measures 1.
Automated insulin delivery (AID) adjusts insulin as glucose changes, which may cushion a miscount. In a two-centre randomized crossover trial led by researchers at McGill University and the Montreal Clinical Research Institute, 30 adults with type 1 diabetes completed 3 weeks on an AID system with each of two meal strategies 5. In one, they counted carbohydrate as usual. In the other, they only classed each meal as low, medium, high or very high in carbohydrate (under 30 g, 30-60 g, 60-90 g or over 90 g), and meal boluses were calculated as if the meal held 15, 35, 65 or 95 g 5.
Time between 3.9 and 10.0 mmol/L (70-180 mg/dL) averaged 74.1% with counting and 70.5% with meal-size categories, and time below 3.9 mmol/L stayed under 1.6% with both 5. The 3.6-point gap works out to about 50 minutes a day, and the trial could not show that the simpler method was as good as counting 5.
The International Society for Pediatric and Adolescent Diabetes (ISPAD) notes in its 2022 nutrition guideline that accurate carb counting has no universal definition, and that accuracy and consistency help post-meal glucose and reduce variability 6.
In the Montreal study, 50 adults kept 3-day food records with their carb estimates while wearing a continuous glucose monitor (CGM), and a dietitian analyzed the same records with nutrition software 1. Across 448 meals averaging 72 g of carbohydrate, estimates differed from the dietitian’s by 15.4 ± 7.8 g per meal, or 20.9 ± 9.7%, and 63% of meals were underestimated 1. The dietitian worked from the written records, so a portion misjudged in the record itself wouldn’t count as an error.
A study that weighed the food found larger errors. At Bern University Hospital in Switzerland, 19 adults with type 1 diabetes each estimated six varied meals from the hospital restaurant, whose carbohydrate content the researchers calculated from weighed portions 7. Their mean absolute error was 27.9 g (SD 38.2), or 54.8% of the meal’s carbohydrate. It grew with meal size, from 18.5 g for small meals to 38.8 g for large ones 7.
In Newcastle, 102 children and teens aged 8.3 to 18.1 years using intensive insulin therapy, and 110 caregivers, each estimated 17 standardized meals containing 8 to 90 g of carbohydrate 8. Of all the estimates, 73% fell within 10-15 g of the true amount 8. Individually labelled foods were estimated correctly most often, and everyday core foods in non-standard portions were missed most often 8. The authors concluded that repeated education is needed to keep estimates accurate 8.
At the SickKids diabetes clinic in Toronto, 140 teens aged 12 to 18 who counted carbs regularly estimated test trays of meals and snacks 9. On the meal trays, 42% were accurate to within 10 g, 44% were off by 10 to 20 g and 14% by more than 20 g 9. Accurate counters scored higher on a carb-counting quiz than those off by more than 20 g, and the researchers called for regular accuracy checks and re-education 9.
Regulators decide how far packaged food may stray from its Nutrition Facts table, and for carbohydrate the rules lean one way.
In Canada, the Canadian Food Inspection Agency (CFIA) checks label values with a laboratory compliance test 10. Inspectors take at least 12 packages at random from a production lot and combine them into three pooled samples of at least four packages each. For carbohydrate, the average of the three samples must be at least 80% of the declared amount 10. There is no fixed upper limit. The food may contain more carbohydrate than declared by an amount consistent with good manufacturing practice, provided the extra poses no health risk and isn’t misleading 10. Calories, fat (including saturated and trans fat), cholesterol, sodium, sugars and sugar alcohols face the opposite limit and must average no more than 120% of the declared amount 10.
The US Food and Drug Administration (FDA) rule follows the same pattern 11. It tests a composite of 12 packages, one from each of 12 randomly chosen shipping cases. For naturally occurring nutrients, total carbohydrate included, the composite must hold at least 80% of the label value, and reasonable excesses are acceptable under current good manufacturing practice. Calories, sugars, total, saturated and trans fat, cholesterol and sodium may not exceed the label by more than 20% 11.
In plain terms: A packaged food must average at least 80% of the carbohydrate on its label, and no fixed number caps how much more it may hold. Compliance is judged on pooled samples, so single packages aren’t tested one by one.
A team at the US National Institute of Diabetes and Digestive and Kidney Diseases in Phoenix tested 24 popular snack foods in the laboratory 12. After accounting for serving weight, calories were 4.3% higher than labelled. In a small subset analyzed for carbohydrate, carbohydrate came in 7.7% above the label, while fat and protein didn’t differ significantly from it 12. On a snack labelled 30 g of carbohydrate, 7.7% is a little over 2 g. The authors judged the products to be well within FDA limits 12. The sample was small and limited to snacks.
A barcode scan looks up the product in a database and returns the values stored there, so at best it reproduces the label, tolerances included. Researchers at Wageningen University & Research in the Netherlands scanned 100 products from the country’s two largest supermarket chains with seven diet apps and compared the results with the package labels 13. The share of products each app identified correctly ranged from 96% (MyFitnessPal) to 5% (SparkPeople). Carbohydrate was listed for 92% of correctly identified products, and the accuracy of nutrient values varied widely between apps and between nutrients 13.
The authors pointed to three weak spots. Many database entries are added by users and can be incomplete or wrong. Manufacturers change recipes, which leaves old values in databases. The same product can also differ in content, and in labelling rules, from one country to another 13. A product bought in Canada may not match an entry created from a US or European package. The study dates from 2018, and app databases change.
GoCARB is a smartphone system designed to estimate the carbohydrate in plated meals for people with type 1 diabetes, studied by researchers at the University of Bern and Bern University Hospital 7, 14, 15. The user takes two photos of the plate with a reference card beside it, and the software separates the foods, recognizes them and calculates their carbohydrate 7, 14.
In the Bern comparison described earlier, GoCARB’s estimates of the same meals missed by 12.3 g on average (SD 9.6), less than half the 27.9 g error of the participants’ own estimates 7. Its automatic separation of foods worked on 75% of meals, and it recognized 85% of individual food items 7.
A second study compared GoCARB with six experienced dietitians, who estimated 54 plated Central European meals, each holding three weighed foods, from photographs 15. The two were level, with mean absolute errors of 14.8 g for GoCARB and 14.9 g for the dietitians. Both missed by more on larger meals, and foods richer in carbohydrate were harder to estimate. GoCARB struggled with rice, pasta and potatoes, and the dietitians with pasta, chips, rice and polenta 15.
A randomized pilot study then looked at glucose. Twenty adults with type 1 diabetes using sensor-augmented pumps practised with GoCARB for two weeks, and then one week of using it was compared with conventional carb estimation 14. With GoCARB, they spent 15.0% of the time above 12 mmol/L (216 mg/dL) compared with 18.2%, and glucose varied slightly less (standard deviation 3.0 vs 3.2 mmol/L). Time below 3.5 mmol/L (63 mg/dL) and the rise in glucose over 3 hours after meals didn’t differ significantly 14. The app suggested a carb amount, and each person decided on their own bolus 14. One week in 20 people is a short test.
At the university hospitals of Caen and Strasbourg in France, trained medical students acting as patients estimated 246 hospital meals with four dedicated carb-counting apps and with ChatGPT-5 16. Dietitians set reference values from weighed food. Mean absolute errors ranged from 13.0 g for the best app to 20.6 g for the least accurate, with ChatGPT-5 at 18.0 g (SD 16.0). Between 50% and 80% of meals came within 20 g of the reference, depending on the tool, including 73% for ChatGPT-5 16.
At the Royal College of Surgeons in Ireland and University College Dublin, researchers gave ChatGPT-4 114 photographs of 38 typical Irish meals, each shown in three portion sizes, with nutrient values worked out from weighed food 17. It identified the foods with 93% precision but underestimated the weight of 76% of meals, and for large portions it estimated 530 g on average when the meals weighed 798 g 17. Carbohydrate came out 6.5% low on average, and an average can mask bigger misses on single meals 17.
In Denmark, researchers at Aalborg University asked ChatGPT-4o to estimate the carbohydrate in 60 fruits and vegetables and 60 composite meals, each photographed with and without a size reference, for 240 images in all 18. Judged against manual carb counts, 93.3% of fruit and vegetable estimates fell within 10 g (95% with the size reference). For composite meals, 46.7% did (43.3% with the size reference) 18. The authors urged caution, since errors on mixed meals could lead to inaccurate insulin doses 18.
None of these three studies measured glucose, and none tested home-cooked meals photographed by people with diabetes in daily life.
| Study | What was tested | Reported accuracy |
|---|---|---|
| Bern, 2016 7 | 19 adults vs GoCARB, 114 meals | Mean miss 27.9 g (people), 12.3 g (GoCARB) |
| Bern, 2018 15 | 6 dietitians vs GoCARB, 54 meals | Mean miss 14.9 g vs 14.8 g |
| France, 2026 16 | 4 apps and ChatGPT-5, 246 meals | Mean miss 13.0-20.6 g; 50-80% within 20 g |
| Ireland, 2025 17 | ChatGPT-4, 114 meal photos | Carbs 6.5% low on average; weight low in 76% |
| Denmark, 2026 18 | ChatGPT-4o, 240 images | Within 10 g: 93% fruit and veg, 47% mixed meals |
Most weak spots come down to what a camera can’t see or measure.
Photo tools can help, as the Bern studies showed 7, 14, 15, and they work best as a draft you check.
These habits concern the carb number itself. Any change to how you dose belongs with your care team.
In a Montreal study, adults with type 1 diabetes misjudged their meals by 15.4 g on average, about 21% of the meal, and 63% of meals were underestimated 1. A Bern study that weighed the food found larger errors, averaging 27.9 g 7.
In controlled studies of children taking mealtime insulin, a 10 g miss on one meal changed little, while a 20 g miss raised the chance of a low or a high 2, 3. The authors concluded estimates should fall within 10 g 3. The margin may differ for adults or larger meals.
Canadian and US rules require packaged food to hold at least 80% of the carbohydrate declared on its label, on average, with no fixed ceiling on extra carbohydrate 10, 11. In a laboratory test of snack foods, carbohydrate came in 7.7% above the label in a small subset 12.
A barcode scan returns the database entry for the product, so at best it reproduces the label. In a Dutch study of seven diet apps, the share of products each app identified correctly ranged from 5% to 96%, and entries could be user-added, incomplete or out of date 13.
In a Bern study, the photo system GoCARB missed by 12.3 g on average, less than half the error of people's own estimates 7. ChatGPT-4o came within 10 g for 93% of fruit and vegetable photos but 47% of mixed meals 18. Treat any photo estimate as a draft to check.
You can check your own counting the way researchers do. Pick a few meals you eat often, weigh or measure them once, and compare the result with your usual estimate. In Glyve, Meal Scan photographs a meal, a nutrition label or a barcode and estimates carbs, protein, fat and fibre. Every estimate is a draft you review and edit before saving, which is the moment to check the portion, the serving size and anything the camera can’t see. The History graph shows meals on the same timeline as glucose, so you can look at what followed meals you weighed and meals you estimated quickly. Glyve shows what happened together; it doesn’t tell you what caused it. Bring what you notice to your care team.
How we write about diabetes. Spotted a mistake? Email info@glyve.ca and we’ll check it against the source.
Glyve keeps glucose, food, insulin, exercise, sleep and stress on one timeline.
For iPhone, iPad and Apple Watch.