We build a photo calorie tracker, so you'd expect us to tell you it's magic. It isn't. Here's the honest version — the numbers most apps in this category quietly leave out of their marketing.
The question everyone asks before trusting a photo app
You point your camera at a plate of pasta, an app says "620 kcal," and your first thought is: says who? Fair. If the number could be off by half, the whole habit is pointless. So, are AI calorie counters accurate? Let's separate what photo AI is genuinely good at from where it still guesses.
What's solved: recognizing the food on the plate
Identifying what you're eating is mostly a solved problem. Modern vision models recognize common dishes correctly around 85–95% of the time. Grilled chicken, a burrito bowl, scrambled eggs — the model knows. Independent reviews of Lose It's "Snap It" feature have put its food identification somewhere around 70%, and reviews of Cal AI find it lands within roughly 10% on simple, fully visible meals. Recognition failures are now the exception, not the rule.
What isn't: portions, hidden oils, and mixed dishes
Portion size is where the real error lives. A photo is a flat 2D image of a 3D plate — the model can't weigh your rice. Portion estimates from images typically run ±15–30%, and studies comparing image-based methods against gold-standard measurements find they underreport intake by around 20% on average.
Three things make it worse:
- Hidden calories. The tablespoon of olive oil in the pan, the butter in the sauce. Invisible in the photo, very visible on the scale. Reviewers testing Cal AI consistently flag missed oils and hidden ingredients.
- Mixed and wrapped dishes. Stews, casseroles, burritos, anything layered. The same Cal AI reviews reported errors of 25% or more on mixed or wrapped dishes — versus ~10% on simple ones.
- Depth ambiguity. A shallow wide bowl and a deep narrow one can photograph almost identically while holding very different amounts.
So the honest error budget looks like this: recognition contributes a little, portions contribute a lot.
The cuisine bias nobody mentions
Most training data for food AI skews heavily Western. A cheeseburger gets a tighter estimate than a plate of menemen, plov, or a home-style Turkish stew — not because those foods are harder, but because the model has seen fewer of them. If you eat outside the Western canon (many of our users do — Kalo ships in English, Turkish, and Russian), expect wider error bars, and correct more often. We're not immune to this; nobody is yet.
Why ±20% is still good enough — if you correct fast and log daily
Here's the part that sounds like a cop-out but is backed by research: consistency beats precision. The "Log Often, Lose More" study (Obesity, 2019) found logging frequency — not accuracy — was the significant predictor of weight loss, and that the time cost of tracking dropped from ~23 minutes a day to ~15 as people got fluent. Other work shows the best adherence marker is simply the number of days you log at least two meals.
A ±20% estimate you actually record every day beats a gram-perfect entry you abandon after two weeks. And errors partially cancel: overestimate lunch, underestimate dinner, and your weekly trend — the number that matters — stays useful.
Two honest caveats. If you need clinical precision (medical conditions, physique competition), a food scale plus a big verified database like MyFitnessPal's is still the better tool — that's a real strength of theirs, along with restaurant coverage, though it now costs $79.99/yr for Premium and its crowdsourced database has its own duplicate-entry problems. And if you want the cheapest big-name premium, Lose It at $39.99/yr is hard to beat, even if its portion estimation tests inconsistently.
How Kalo handles uncertainty: see it, fix it in two taps
Since portion error is unavoidable, the design question becomes: what does the app do about it?
Kalo shows you its confidence instead of pretending certainty — a shaky estimate looks shaky. And correcting one is a two-tap act on the result card itself, not a five-screen dig through an edit menu. Snap, glance, nudge the portion, done. The faster correction is, the more often it actually happens, and the tighter your real-world accuracy gets.
The other half is why we built Kalo around friends. You can add up to 5 friends, see everyone's day side by side, react and nudge, and build streaks together — because the research says a partner changing with you is the biggest lever there is (in one JAMA Internal Medicine study of ~3,700 couples, smokers' chances of quitting jumped from about 8% to nearly 50% when their partner quit at the same time — and the same pattern held for exercise and weight loss). Accuracy keeps a number honest; company keeps the habit alive. One friend can be starred as your partner and shares your Premium.
If you want a tracker that's honest about its error bars and easy to correct — and more fun with your people in it — Kalo is on iOS and Android.