Can ChatGPT count calories from a photo? An honest look

Short answer: yes, it can guess, and research shows how good the guess is

Can ChatGPT count calories from a photo? Yes. Upload a meal picture, ask for calories, and it will name the foods and produce an estimate almost every time. The catch is in the word estimate: published evaluations show it names foods well but gets portions, and therefore calories, wrong often enough to matter.

A 2025 study in the journal Nutrients tested ChatGPT-4 on 114 meal photographs drawn from national dietary survey data (O'Hara and colleagues). It was good at naming what was on the plate. Its weight estimates agreed well with reality for small meals but poorly for medium and large ones, and it underestimated 11 of the 16 nutrients it was asked about.

In practice, an estimate that runs low on bigger meals is the wrong kind of error if you're trying to eat less: the missing calories never show up in your total.

Why a chatbot struggles with food photos

These aren't random glitches. They follow from how a general-purpose chatbot works.

• No real portion step. A single photo has no depth and no reliable size reference. A chatbot can't tell a 25 cm plate from a 30 cm one, or a small mound of rice from a big one piled the same way.
• No checkable nutrition source. The answer isn't tied to a nutrition database entry you can see and correct, which is part of why the same photo can come back with different numbers if you ask twice.
• Hidden calories are invisible. Cooking oil, butter, sugar dissolved in a sauce, dressing already tossed through a salad: none of it shows up in the pixels.
• It sounds confident either way. The answer arrives in the same assured tone whether it's close or far off, so you get no signal about when to double-check.

Where ChatGPT is genuinely fine

To be fair, there are situations where asking ChatGPT for calories from a photo works well enough:

• Ballpark sanity checks. "Is this closer to 400 or 900 calories?" is a question it can usually settle.
• Single, obvious foods. A banana, an apple or a plain grilled chicken breast: simple foods with standard sizes are the easiest case.
• When you type in the details. Studies have found estimates get much better when the food's weight is supplied (one 2025 evaluation of three AI models). At that point, though, you're doing much of the work the photo was supposed to save you.
• General nutrition questions. Explaining why fiber matters or what's typically in a burrito is squarely in its comfort zone.

What a dedicated photo scanner does differently

Apps built for counting calories from a picture work on those weak points directly. Foodify's plate scan lists each food on the plate separately, estimates a portion for each, and returns calories plus protein, carbs and fat for each food, not one lump-sum guess for the whole plate.

Just as important, every result is editable before you save it. No camera-based tool sees hidden oil or knows your exact serving (photo estimates have real limits across the board), so a good scanner treats its output as a first draft you can correct in two taps. For packaged foods, a barcode scan skips estimation and pulls the nutrition facts from the Open Food Facts database.

The workflow problem: a chat doesn't track anything

Suppose every estimate were spot on. You would still be missing the thing that makes calorie counting work: the log. There's no running daily total, no protein, carb and fat breakdown against your targets, no weekly trend and no history to look back on. Each answer sits in a chat thread, and connecting yesterday's lunch to today's dinner means scrolling, copying and adding up by hand. A chat history is not a food diary.

This is the quiet reason "just ask the chatbot" fades after a week. Calorie counting pays off over time: seeing that you eat more on weekends, that protein dips on busy days, that the trend line is finally moving. A chatbot gives you disconnected snapshots; an app gives you the whole film.

You don't have to give up the chat, either. Foodify includes Foodi, an AI coach for food questions: ask anything in plain words or send a photo, and the answer is about your own day. Add water, weight and body tracking, plus activity read from the Apple Health app to balance calories in and out, and the gap between a chatbot and a tracker gets wide. If you're comparing dedicated apps, our guide to questions worth asking any AI calorie app walks through what to check.

Bottom line: ChatGPT can count calories from a photo the way a well-read friend can: a reasonable guess, delivered confidently, remembered by no one. For an occasional sanity check, that's fine. For tracking what you eat, a scanner with portion estimates, editable results and a real log is the tool for the job.

FAQ