AI nutrition tracking needs a major upgrade

For anyone serious about fitness or health goals, tracking nutrition with current AI tools often feels more like a chore than a helpful innovation. The technology promises to simplify logging meals by analyzing photos, but in practice, it frequently misidentifies foods and miscalculates portions, leaving users to manually correct errors. This defeats the purpose of saving time and can make the entire process more frustrating than traditional methods.

Take my own experience. Before a long run, I typically eat two dark chocolate Kodiak protein waffles with a tablespoon of peanut butter and a drizzle of honey, alongside iced coffee with a splash of soy milk. After several training cycles, I’ve dialed in that this provides about 355 calories, 16 grams of protein, 28 grams of carbs, and 17 grams of fat, just enough to fuel my workout without a post-run crash. Recently, my strength training app, Ladder, introduced an AI nutrition feature that claimed to make macro counting effortless. All I had to do was snap a picture. Instead of accurate data, the AI estimated my breakfast at 780 calories, 20 grams of protein, 92 grams of carbs, and 39 grams of fat. Even after I manually edited the entry to specify brands and exact amounts, it still produced wildly incorrect numbers. Situations like this are exactly why I’ve largely abandoned tracking calories and macros.

Let’s be honest: logging everything you eat is tedious. Standard apps let you search databases or scan barcodes, which works fine for packaged or whole foods. The system falls apart, though, when you eat at restaurants or cook meals at home. Many restaurants list calorie counts without macro breakdowns, and while some apps let you import online recipes, that’s little help if you’re improvising dinner or swapping ingredients. To log with any precision, you practically need to weigh every component, avoid dining out, and stick to a repetitive menu. That routine gets old quickly.

Research does show that keeping a food diary or using digital tracking tools can support weight management and muscle gain. That’s probably why so many health and fitness apps are now integrating AI to reduce the drudgery. The options seem endless. Oura rolled out a chatbot advisor that lets you describe meals or take photos, then provides macro estimates, notes on processing level, and potential health impacts. If you use a Dexcom continuous glucose monitor, you can sync that data and see how specific foods affect your blood sugar. The January app also uses photos and user demographics to predict a meal’s likely impact on glucose. MyFitnessPal has a ScanMeal feature for photo-based calorie and macro estimates. My TikTok feed even pushes a gamified app where an AI raccoon “eats” the meals you photograph while logging the data. Ladder’s AI accepts photos, voice descriptions, or typed entries. Despite different interfaces, the core promise is the same: snap a picture and let artificial intelligence handle the rest.

Sadly, AI remains mediocre at recognizing foods from images. Oura Advisor repeatedly identified my matcha protein shakes as green smoothies. January recognized chicken in one meal but confused barbecue sauce with teriyaki and missed the mushrooms entirely. When Ladder botched my breakfast log, it assumed I’d eaten two seven-inch waffles instead of four-inch ones, doubled the peanut butter, swapped honey for syrup, and added cream and sugar to my coffee, which I never use. None of these systems noticed when I made healthier substitutions, like mixing edamame and quinoa into brown rice for extra nutrients. Oura’s AI labeled that combination as mashed potatoes and white rice. Ethnic dishes are especially hit-or-miss. Ladder logged my dal makhani curry with basmati rice and peas as chicken soup. Sometimes it correctly identifies tteokbokki, the Korean rice cakes in spicy sauce; other times, it calls them rigatoni in tomato sauce.

You can edit these AI-generated entries, of course. But doing so undermines the whole goal of simplifying a boring task. Instead of saving time searching for foods to log, you waste minutes fixing the AI’s mistakes. You’re just trading one annoyance for another.

After some reflection, maybe using AI to simplify food logging is solving the wrong problem. Current image recognition can distinguish broad categories, a banana from an apple, but it can’t identify ravioli fillings or accurately gauge portion sizes. If precision matters, you’ll always need to supervise it. More fundamentally, this approach ignores the real challenge: dietary changes are difficult not because we lack information, but because applying that knowledge consistently requires shifting behaviors and emotional habits. AI can suggest adjustments, but you’re still the one who has to implement them.

The true purpose of food logging isn’t hitting arbitrary calorie or macro targets. It’s about building awareness, learning your eating patterns, recognizing areas for improvement, and practicing mindfulness, even when you treat yourself. Once you internalize those lessons, you ideally stop logging. You might resume temporarily if your goals or health changes, but it’s not meant to be a lifelong habit. The aim is to reach a point where you trust your own judgment about what and when to eat.

App developers, however, have little incentive for you to quit. A “successful” food logging app keeps you engaged indefinitely. Instead of attributing progress to your own hard-earned understanding, you credit the tool. You might worry that without constant tracking, you’ll backslide. Or if you’re struggling, you hope AI will make a difficult process easier, though it often doesn’t.

There is potential in the concept of photographing your meal and receiving a useful insight from AI. I’m just not sure what that insight should be. Maybe it would help if AI acknowledged my home-cooked meal as a nutritional win, pointed out a 15 percent increase in glazed donuts over the past month to prompt reflection on stress eating, or gently suggested, “You’ve eaten a lot of baked chicken breasts lately, maybe treat yourself to some white rice.” What I do know is that AI shouldn’t force me to photograph my breakfast and then spend the next quarter-hour arguing with it to correctly identify what I actually ate.

(Source: The Verge)