Can ChatGPT count calories from a photo? What the studies found

Ask ChatGPT how many calories are in a photo of your dinner, and it will give you a number in seconds, often with a confident breakdown. That makes it tempting as a free calorie counter. Researchers have now tested exactly this, and their results are useful whether you use ChatGPT, another AI model or a calorie app.

What the studies found

General AI models miss by about a third

The most direct test comes from a 2025 study that gave three AI models 52 standardized photos of single foods and meals whose weights were known. ChatGPT-4o and Claude 3.5 Sonnet missed the calories by 35.8% on average; Gemini 1.5 Pro by 64.2%. All three under-estimated more as portions got bigger: errors were about 20 to 30% lower for small portions than for large ones.

Good at naming foods, weaker at amounts

A 2025 study of ChatGPT-4 on 114 meal photos, 38 meals each shown at three portion sizes, found it identified foods well, with 93% precision. Amounts were the problem: large meals that weighed 798 g on average were estimated at 530 g, and medium meals were under-estimated too, while small meals were close. Its estimates differed significantly from the true values for 10 of the 16 nutrients studied. One caveat in ChatGPT’s favor: on calories specifically, its median estimate was close to the true median in that study.

Details make a big difference

The strongest evidence that you can help comes from a 2025 study of ChatGPT-5 on 195 dishes. From the photo alone, its calorie estimates were off by 30.5% on average (123 kcal). Told the type and amount of fat, the sweetener, the fat content of dairy and the type of meat, the error fell to 24.4% (92 kcal). Given the ingredient list with amounts, it fell to 13.9% (53 kcal). A small 2026 pilot of ChatGPT-4o on nine dishes points the same way: errors reached 54.4% for calories and 76.5% for fat on complex dishes with hidden fat, and adding details made its estimates far more consistent. With only nine dishes, treat the 54% as a worst case, not an average.

Other findings worth knowing

• People still do better. In a 2026 test of 40 AI models, professional nutritionists estimated nutrients more accurately than every model, and photos from several angles didn’t help the models.
• AI can invent things. Researchers who built a dietary tool on GPT-4o found plain GPT-4o recognized 59% of foods from photos alone and warned about “hallucinations (fabricated items, units, or quantities)” (AJCN, 2025).
• Typed foods are easier than photos. When AI models were given food names as text instead of photos, their energy estimates agreed well with nutrient databases, with mixed dishes and condiments varying most (Journal of Nutrition, 2026).

Why ChatGPT misses on photos

The same three things trip up ChatGPT and photo calorie apps:

• Portion size. A photo is flat, so the depth of a bowl of pasta or rice is hard to judge, which is why big portions get under-counted.
• Hidden fat. Oil soaked into vegetables or butter melted into a sauce barely shows. One tablespoon of olive oil is 119 kcal, according to USDA FoodData Central.
• Mixed dishes. In a curry, a burrito or a casserole, the AI has to guess the recipe as well as the amount.

Dedicated apps aren’t immune: in a preliminary 2026 NIH test, the photo features of MyFitnessPal, Lose It!, Cal AI and Appediet under-counted weighed meals by 252 to 345 kcal on average, about a third. Our review of photo calorie apps covers that study and each app’s accuracy claims.

ChatGPT vs a calorie app

The estimate itself isn’t magic in either case: many calorie apps, Calorix included, use large AI models to read photos. The difference is what surrounds the estimate. In Calorix, a photo comes back as a list of ingredients with grams and a confidence badge you can correct, barcodes and nutrition labels cover packaged food, and everything lands in a log with your targets.

How to get better estimates from ChatGPT, or any AI

The research says the biggest gains come from telling the AI what the photo can’t show. A prompt like this one works better than a photo alone:

Estimate the calories, protein, carbs and fat in this photo. It’s about 180 g of cooked rice, 150 g of grilled chicken thigh and roasted vegetables cooked in about 1 tablespoon of olive oil. List each ingredient with the amount and calories you assumed, then the total.

• Give weights or sizes when you know them: grams from a scale, “a full dinner plate”, “half a cup”.
• Name the cooking fat and how much: oil, butter, ghee, cream.
• Mention sauces, dressings and drinks, which photos miss.
• Ask for a per-ingredient breakdown with its assumptions, so you can see where it guessed.
• Shoot at a 45-degree angle with the whole plate in frame, rather than straight down.
• Use the label for packaged food. The printed values beat any estimate.

The bottom line

ChatGPT can count calories from a photo well enough for a rough idea, and much better when you give it the amounts and the cooking fat. For tracking day after day, a calorie app adds the parts ChatGPT isn’t built for: a log with targets, barcodes, quick corrections and trends. Whichever you use, treat a photo estimate as a first draft. If you want a target to log against, start with our TDEE calculator.