The promise is seductive: Snap a photo of your meal, and artificial intelligence will instantly tell you exactly how many calories you’re consuming. No more tedious manual logging, no more guessing at portion sizes, no more human error. Apps like Cal AI, Lose It!, and MyFitnessPal’s new photo features claim to revolutionize calorie tracking by letting your smartphone’s camera do the heavy lifting.

But as someone with a long, complicated history with calorie counting—and admittedly, a somewhat cursed expertise in it—I can tell you that counting calories with a photo is exactly as stupid as it sounds.

How AI-powered calorie counting is supposed to work

Calorie counting apps promise to solve what developers claim is the biggest problem with calorie tracking: human error. The pitch is compelling—why spend time searching databases and measuring portions when your phone can instantly analyze your plate?

Apps like Cal AI or SnapCalorie AI use visual cues like color, texture, and relative size to make educated guesses about what you’re eating and how much of it there is.

They claim that AI methods can solve the pesky problem of human accuracy in calorie estimation—which, to be fair, is easy to get wrong. Cal AI markets itself as one of the more sophisticated options in this space, so I decided to see for myself. The app was free for the first three days, then $29.99/year.

The setup process is straightforward: Download the app, create an account, input basic demographic information, and set your goals. Here’s where I encountered my first red flag. The app cheerfully informed me that “losing 10 lbs is a realistic target”—except that losing 10 pounds would actually push me into underweight BMI territory. This kind of blanket statement reveals a concerning lack of nuance about individual health needs.

Cal AI’s photo logging process follows these steps:

1. Take a clear photo of your food, ideally against a plain background.

2. Make sure all ingredients are visible and well-lit.

3. Include a reference object (like a coin or your hand) for scale.

4. Upload the image and wait for AI analysis.

5. Review and correct the app’s identifications and portion estimates.

6. Save the entry to your daily log.

The app provides detailed tips for getting better results: Use natural lighting, avoid shadows, keep the camera parallel to the plate, and ensure no ingredients are hidden. These guidelines sound reasonable in theory, but they hint at the fundamental challenge these apps face—the complexity of real-world eating.

The reality is wildly disappointing

I started my testing with something simple: a Pink Lady apple weighing 222 grams. Surely this would be an easy win for the AI—apples are among the most photographed foods on earth, with a distinctive shape and color that should be instantly recognizable.

Cal AI confidently identified my apple as tikka masala.

I gave it another chance, this time photographing the apple alongside its barcode and sitting on a kitchen scale displaying its exact weight. The app did recognize it as an apple this time, but estimated it at 80 calories when the actual count should have been closer to 120. That’s a 33% underestimation—not exactly the precision you’d want if you’re trying to track intake accurately.

The real test came with a more complex meal: my current meal-prepped lunch of fried tofu, onions, cucumbers, tomatoes, feta cheese, and chickpeas, all generously dressed with an oil-based homemade vinaigrette. This is the kind of mixed dish that presumably showcases AI’s advantage over manual logging—no need to search for individual ingredients or estimate their quantities.

The results were a masterclass in algorithmic overconfidence. The app identified the golden-brown fried tofu as croutons, which I had to manually correct. It did a reasonably good job recognizing the vegetables and feta, but completely whiffed on the oil content. Despite the salad being visibly glistening with dressing, the app estimated the entire dish at 450 calories.

This estimate was laughably low. A single can of chickpeas contains about 400 calories, and my portion included roughly that amount plus significant quantities of feta cheese and several tablespoons of olive-oil-based dressing. A realistic calorie count for this meal would have been closer to 800 to 900 calories.

The app’s portion estimation proved even more problematic than its ingredient identification. When I photographed a smaller serving—less than a quarter of the original salad—Cal AI estimated it at 250 calories. According to the app’s own logic, less than 25% of the meal somehow contained more than 55% of its calories. The math simply doesn’t work.

This highlights a fundamental limitation of photo-based calorie counting: Cameras capture two-dimensional images of three-dimensional objects. Without consistent reference points or sophisticated depth analysis, estimating volume from photos remains largely guesswork. Even humans struggle with this task, which is why nutrition professionals typically recommend weighing foods for accuracy.

To get a fuller picture of the AI calorie counting landscape, I also tested two other popular apps: SnapCalorie and Calorie Mama.

SnapCalorie: better numbers, same problems

SnapCalorie did immediately assuage some skepticism by suggesting a much more reasonable daily calorie target of 1,900 calories, compared to Cal AI’s problematic weight loss messaging. However, this accuracy comes at a steep price—$79.99 per year after just a one-week free trial, making it the most expensive option I tested.

The app does offer one interesting feature: an “add note” function that lets you provide additional context about ingredients the camera can’t see. In theory, this addresses one of the fundamental limitations of photo-based tracking.

When I tested SnapCalorie with the same Pink Lady apple, it performed much better than Cal AI, estimating 115 calories. But the Greek salad test revealed familiar problems. SnapCalorie’s initial estimate was an absurdly low 257 calories. When I photographed a smaller portion—the same sub-quarter serving that had stumped Cal AI—SnapCalorie estimated 184 calories. The math still didn’t work; this smaller portion should have been roughly 25% of the larger serving, not 70%.

Determined to give the app a fair shot, I used the note feature to manually specify “full container of tofu, feta, chickpeas and olive oil.” With this human intervention, SnapCalorie bumped its estimate to 761 calories—much more reasonable and accurate, though still on the low side.

But this raises the obvious question: If I need to manually input detailed ingredient information to get accurate results, what exactly is the photo accomplishing? I’m essentially doing the work of traditional calorie counting while going through the motions of taking pictures.