# AI Calorie Counter Accuracy: What the Studies Show > AI photo calorie counters name foods well but often misjudge portions, calories and fat. Here is what six PubMed studies found, what the app makers say about their own estimates, why errors happen and how to log more reliably. *Author: Trainera Team | Published: 2026-10-17 | Reading time: 9 min* ## How accurate are AI photo calorie counters? AI photo calorie counters are good at naming the food on your plate and much weaker at judging how much of it there is, so the calorie number can be well off, especially on large, mixed or oily meals. In the six studies below, apps and AI models that were tested on recognition named most foods correctly, while calorie or nutrient estimates were more than 10% off in two of them, portion estimates were poor for medium and large meals in one, and fat was overestimated in another. Treat a photo estimate as a first draft that you check, not as a measurement. Below: the evidence, what app makers say, why errors happen and how to log better. If you are picking an app rather than checking the science, our roundup of the [best calorie counter apps with AI food photo](/blogs/best-calorie-counter-app-ai-photo) compares them. Every source here was checked October 2026. ## What the research found Recognition is strong, portion and calorie estimates are not, and the misses get bigger as meals get bigger. All six abstracts are on PubMed. They test different things, so read each result on its own terms, not as one accuracy score for "AI". | Study | What was tested | Main finding (from the abstract) | | -------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ | | [Li et al., Nutrients 2024](https://pubmed.ncbi.nlm.nih.gov/39125452/) | Seven apps with AI food image recognition, plus manual logging apps | MyFitnessPal and Fastic recognized foods best (97% and 92%), but "automatic energy estimations from AI-enabled food image recognition were inaccurate" | | [O'Hara et al., Nutrients 2025](https://pubmed.ncbi.nlm.nih.gov/40004936/) | ChatGPT on 114 meal photos | 93.0% precision for identifying foods; "poor agreement" on weight for medium and large meals; more than 10% off for 13 nutrients | | [Isobe et al., Nutrients 2026](https://pubmed.ncbi.nlm.nih.gov/41901141/) | 10 AI models vs 10 registered dietitians on 15 standardized hospital meals | Top models were often within 10% for energy and carbohydrates; all models overestimated fat, with a mean bias above +20% | | [Rodríguez-Jiménez et al., Nutrients 2025](https://pubmed.ncbi.nlm.nih.gov/41305663/) | ChatGPT-5 on 195 dishes with more or less context | Accuracy improved when a description and then ingredient amounts were added to the photo, and dropped with no photo at all | | [Lozano et al., Current Developments in Nutrition 2023](https://pubmed.ncbi.nlm.nih.gov/38026571/) | An AI app (Openfit) on covertly weighed meals, 24 adults | Energy error of 43% (automated) and 33% (semi-automated) with drinks included; "larger error for higher energy meals" | | [Serra et al., Frontiers in Nutrition 2023](https://pubmed.ncbi.nlm.nih.gov/37810925/) | A food recognition app (SNAQ) vs doubly labeled water in 30 women | Average bias of -329.6 kcal a day, closer than a 24-hour recall (-543.0), but no relationship between individual estimates and the reference | ### Naming the food is not counting the calories The high percentages you see quoted are usually recognition scores. In the Li study, 97% and 92% describe how often the apps named the food correctly, not how close the calories were. The ChatGPT study tells the same story: 93.0% for identifying foods, yet it "performed poorly for estimating weights of medium and large portion sizes", and it underestimated 11 of the nutrients it was checked on. Getting "chicken, rice, broccoli" right is the easy part. The calories depend on the grams of each. ### Bigger meals, bigger misses Two studies point the same way. The Openfit pilot found larger errors for higher energy meals, and the ChatGPT study found poor weight agreement for medium and large meals. In practice, a snack that is a bit off costs you little; a big dinner plate off by the same share can swing your day. ### Fat is the weak spot The hospital meal study used standardized meals, the most controlled setting in this list. Even there, accuracy for protein and fat was "significantly lower" than for energy and carbohydrates, and every AI model overestimated fat by more than 20% on average. If you track macros, the fat line on a photo log deserves the most suspicion. ### A good average can hide bad single days Doubly labeled water is a lab method that measures how much energy a person burns over several days, which researchers use as a reference for intake. Against it, the SNAQ app did better on average than a 24-hour recall interview. The same abstract reports no relationship between the app's estimates and the reference for individual people, and notes that the app company built the study version for payment. A group average that looks close does not tell you your own numbers are close. ### More context helps The ChatGPT-5 study is the most useful one for everyday logging. Of the three photo setups, photo only was the least accurate, photo plus a description was better, and photo plus ingredient amounts was best. Typing "cooked in olive oil, about 150 g rice" next to the picture is worth doing. ## What the app makers say about their own estimates None of the vendors we checked calls a photo estimate final, and most tell you to review it before saving. | App | What the vendor says (exact wording) | Source | | ------------ | ------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------- | | Cal AI | "CalAI is about 80% accurate. No food tracking app is perfect, and calories and nutrition values can vary." | [Cal AI FAQ](https://www.calai.app/faq) | | MyFitnessPal | "Meal Scan gives you a starting point, not a final answer, so you always have the last word on what goes in your diary." | [MyFitnessPal help](https://support.myfitnesspal.com/hc/en-us/articles/360045761612-How-to-use-the-Meal-Scanner) | | Yazio | "The AI recognition only provides a rough estimate and cannot guarantee 100% accurate results." | [Yazio help](https://help.yazio.com/hc/en-us/articles/39137901903889-What-is-the-AI-Calorie-Tracking-Feature-and-how-does-it-work) | | Cronometer | "Always check if your foods have been matched correctly before adding to your diary for the most accurate logging." | [Cronometer support](https://support.cronometer.com/hc/en-us/articles/39013533811092-Mobile-Photo-Logging) | | MacroFactor | "AI is not perfect and can make mistakes, and we give users the control to edit and modify its results." | [MacroFactor comparison page](https://macrofactor.com/macrofactor-vs-cal-ai/) | | Lose It! | "We'll do our best to suggest the best match for the food(s) in your photo, but you can always make adjustments." | [Lose It! help](https://help.loseit.com/hc/en-us/articles/47771695186580-How-to-Use-Snap-It) | Two of these deserve a note. Cal AI's "about 80%" comes with no published method, so you cannot tell whether it means food recognition, calories per meal or something else. MacroFactor says that in its own side-by-side test with weighed food it came closer than Cal AI, but that is a vendor test with no data on the page, not an independent study. MyFitnessPal's help page adds a practical tip worth copying in any app: "If your meal has several items, scanning one dish at a time usually gives you more accurate suggestions." ## Why photo estimates go wrong A photo shows the surface of a meal, while calories depend on weight and on ingredients the camera cannot see. * **Hidden fat.** Oil soaked into rice or vegetables, butter melted into a sauce and dressing at the bottom of a bowl look like nothing on camera. Fat is the most energy-dense part of food, so a small miss here moves the total a lot. * **Portion depth.** A picture is flat. A shallow plate of pasta and a deep bowl of it can look similar from above. Some apps try to measure volume (Cal AI says your "phone's depth sensor calculates food volume"), but the studies above still found portion size to be the main weakness. * **Mixed dishes.** Stews, curries, casseroles, burritos and sandwiches hide their ingredients. The model has to guess the recipe as well as the amount. * **Look-alikes.** Whole and skim milk in coffee, regular and light mayonnaise, sugary and diet drinks look the same in a photo. * **Drinks.** In the Openfit study the error changed noticeably depending on whether beverages were counted, so a glass on the side matters. * **Local recipes.** The same dish name can mean very different recipes from one country or one kitchen to another. Our guide to [tracking Middle Eastern foods](/blogs/calorie-counter-middle-eastern-foods) shows how much this varies. | Meal type | Typical problem for a photo | What to do | | ------------------------------------- | -------------------------------------------------- | ---------------------------------------------------------------------------- | | Packaged food | The photo guesses what the label already tells you | Use the label or barcode instead of a photo | | Simple plate (meat, rice, vegetables) | Portion of rice or pasta, cooking oil | Correct the grams of the starch and add the oil | | Salad with dressing | Dressing and cheese hidden or underestimated | Log the dressing separately | | Stew, curry, casserole | Unknown recipe and fat content | Describe the main ingredients, or log your own recipe once and reuse it | | Restaurant meal | Recipe, cooking fat and portion are all unknown | Check the restaurant's published nutrition if it has one, otherwise round up | | Coffee and drinks | Milk type, sugar, syrups | Add them as separate items | ## How to get better numbers from any photo tracker Use the photo to name the foods, then spend a few seconds correcting the portions and adding what the camera cannot see. 1. **Check every item before you save.** Delete foods that are not there, rename ones that are wrong and change the cooking method if the app guessed fried instead of grilled. 2. **Fix the portion of the biggest item.** On most plates that is the starch or the protein. If you cooked it, you often know the amount from the packet. 3. **Weigh calorie-dense foods sometimes.** Oil, butter, nut butter, nuts, cheese, granola and dry rice or pasta are where small errors add up. Weighing them now and then trains your eye so your corrections get better. 4. **Add the hidden extras.** Cooking oil, sauces, dressing, sugar in coffee. 5. **Scan one dish at a time** when the plate has several separate foods. 6. **Add words.** If the app accepts a description, type the cooking fat and rough amounts. 7. **Use labels for packaged food.** A barcode or label beats any photo guess. 8. **Judge by the trend.** A single meal can be off. Compare your logged average with your weight trend over a few weeks and adjust from there. Our [calorie deficit guide](/blogs/calorie-deficit-guide-fat-loss) explains how. A common mistake is the opposite one: spending ten minutes perfecting a side salad while the oil in the pan never gets logged. Put your effort where the calories are. ## How accurate does your logging need to be? It depends on your goal: rough is fine for awareness, and the closer you are to a strict target, the more you should weigh. If you are learning what a normal portion looks like or trying to eat a bit more protein, a corrected photo log is plenty. If you are in a tight fat-loss phase or aiming at precise [macro targets](/blogs/how-calculate-macros-fitness-goal-guide), weigh the calorie-dense items and use the photo for the rest. Some situations need more than an app. If you count carbohydrates for medication, have a medical condition, are pregnant or take GLP-1 medication, talk to your doctor before relying on photo estimates. If tracking food makes you anxious, or you have a history of disordered eating, talk to your doctor first, and know that not tracking is also a valid choice. ## Where Trainera's food photo scan fits Trainera's food photo scan gives you an AI estimate that you can edit before logging, and it is part of the Premium plan at $6.99 a month in the US (prices differ by country on [trainera.fit/pricing](/pricing)). When you take a photo, an AI vision model names the foods and estimates the grams, calories and macros of each item. You can correct any item before it goes into your log, which is exactly the habit the studies above support. Premium allows up to 20 scans a day. It is an estimate, like every tool on this page, and it does not replace weighing food when precision matters. What Trainera adds is the training side in the same app: Premium also includes 10 AI training plans, 10 AI nutrition plans and 10 AI workouts a month, AI Nutrition Chat and TRAI, the in-app assistant, with 90 requests a month. The Free plan, with no card, includes the diet tracker with calories and macros, food search and typing in a barcode number. Clients coached by a trainer on Pro, Business or Enterprise get the scan through their coach. If photo logging is the only thing you want, a dedicated tracker may suit you better. Our [Cal AI vs MyFitnessPal](/blogs/cal-ai-vs-myfitnesspal) comparison and the list of [Cal AI alternatives](/blogs/cal-ai-alternatives) cover those apps. Meal photos are personal data too, so it is worth reading [how fitness apps handle your data](/blogs/ai-fitness-app-privacy) before you pick one. ## Your checklist for more reliable photo logs * Trust the food names more than the grams. * Correct the biggest item and add the cooking fat on every meal. * Be most skeptical of the fat number. * Scan dishes one at a time and add a short description. * Weigh calorie-dense foods now and then to recalibrate. * Judge your intake by the weekly trend, not one meal. _Want an AI estimate of your plate next to an AI training plan? [See how Trainera's AI fitness coach works](/ai-fitness-coach) or [start free](/auth/register?role=client)._ ## FAQ ### Are AI calorie counter apps accurate? They are good at recognizing foods but less reliable at estimating portions and calories. Studies on PubMed found high food recognition rates, while calorie or nutrient estimates were more than 10% off in two of them and fat was overestimated in another. Check portions and add hidden oil before you save a photo log. ### Is Cal AI really 80% accurate? Cal AI says it is "about 80% accurate" on its FAQ, but it publishes no method or definition, so you cannot tell what the 80% measures. Treat any photo estimate, from any app, as a starting point to correct. ### Is it better to weigh food or take a photo? Weighing is more accurate, especially for calorie-dense foods like oil, nuts, cheese and dry rice or pasta. A practical middle ground is to photo log most meals, correct the portions, and weigh the calorie-dense items now and then to keep your estimates honest. ### Do AI food scanners overestimate or underestimate calories? Both happen. In one 2026 study every AI model overestimated fat by more than 20% on average, while a 2025 ChatGPT study found underestimates for 11 nutrients and poor portion estimates on medium and large meals. Errors tend to be larger on bigger meals. ### Can ChatGPT count calories from a photo? It can name foods well, but studies found weak portion and nutrient estimates from a photo alone. Accuracy improved when a description and ingredient amounts were added, so type what the photo cannot show, such as cooking oil and rough grams. --- Source: https://trainera.fit/blogs/ai-calorie-counter-accuracy