Ask anyone who has used more than one AI calorie-tracking app why the numbers do not match between apps, and the honest answer is almost never about food identification. It is about portion size.
What the error actually looks like
User reports and independent reviews of Cal AI, one of the category's largest apps, have described portion estimates off by 20 to 40 percent, particularly on calorie-dense foods where a small error in estimated weight becomes a large error in estimated calories. A 20 percent underestimate on a 150-gram serving of pasta is a rounding error. The same 20 percent underestimate on a tablespoon of olive oil, at roughly 120 calories per tablespoon, is the difference between a meal that fits a deficit and one that does not. That specific number is one app's reported experience, not a formal benchmark across the category, but the underlying cause is not specific to that app.
The underlying cause is physics, not implementation quality: a single 2D photo does not contain true depth information, so every app in this category is estimating a 3D quantity, volume, from a 2D signal, then compounding that estimate through a density assumption and a database lookup. Each step can be individually reasonable and still produce a meaningfully wrong final number.
Why this matters more than most marketing admits
Most product pages in this category describe accuracy as a single number, "95% accurate," without specifying what that number is measuring. Food identification accuracy (did the model correctly say "chicken breast") and portion accuracy (did it correctly say "180 grams") are different metrics with very different real-world error rates, and conflating them is one of the most common ways this category overstates itself.
The honest framing: identification is close to solved. Portion estimation is a harder, unsolved problem across the category, and treating it as already solved is why so many users of these apps eventually stop trusting the numbers.
How Arcal approaches it
Portion estimation, not food identification, is the exact problem Arcal's camera system is built around: the wide and ultrawide lenses run together for genuine spatial context instead of a single flat photo, calibrated against real, weighed meals rather than a raw model guess. Join the waitlist for early access.