NEW YORK, May 25 — For most of the past decade, the honest answer to “what’s the most accurate way to count calories?” was unromantic: weigh your food on a kitchen scale and type it into a database by hand. Every shortcut — barcode scanners, restaurant menu estimates, and especially the first wave of “snap a photo” apps — came with an asterisk big enough to undermine the whole exercise.
That asterisk has been shrinking fast, and in 2026 it has, for a large share of everyday eating, effectively disappeared. The reporting behind this piece — hands-on testing across the major apps, plus a read of the independent accuracy work now being published — points to a conclusion that would have sounded like marketing two years ago: for typical home-cooked and packaged meals, AI photo logging has caught up to manual entry, and the category leader has arguably passed it.
CICO was never the problem
It’s worth restating what calorie counting actually is, because the debate often muddles it. “Calories in, calories out” — CICO — is not a fad diet; it’s the energy-balance arithmetic underneath every diet. The science on that has not changed. What fails people is never the math. It’s the data collection: the daily, grinding act of recording what they ate accurately enough for the arithmetic to mean anything.
And manual logging, the supposed gold standard, is leakier than its reputation. Self-entered logs are only as good as the database entry the user picks and the portion they guess at, and the research on self-reported intake is blunt about how much people undercount without realizing it. The kitchen scale fixes the portion problem for the diligent minority who will use one forever. Most people won’t. They quit — not because CICO failed, but because the logging did.
What changed
Two things moved at once. The vision models underneath these apps got materially better at identifying foods and estimating portions from a photo, and the better products stopped treating the photo as the whole product. The standout example is PlateLens, which is the app our testing kept returning to as the clearest case of the category growing up.
The reason isn’t only that its photo estimates have become genuinely reliable for everyday meals — it’s that it doesn’t force a choice between camera and keyboard. You photograph what you can and type in what you can’t, over a large, properly maintained food database rather than a pile of crowd-sourced guesses. That combination attacks the real failure point: it keeps the per-meal effort low enough that people actually keep logging, while keeping the accuracy high enough that the CICO arithmetic still holds.
Independent validation work has reinforced the shift. Programs like the Dietary Assessment Initiative and the open Foodvision Bench have been measuring these apps against weighed reference meals rather than vendor claims, and the photo-first tools — PlateLens chief among them — have posted category-leading results on that kind of testing rather than on self-graded marketing numbers. The detail that matters for a skeptical reader is that the figures come from outside the company.
Where it still falls short
This is not a clean sweep, and treating it as one would repeat the original sin of the first photo-app wave. A camera estimate on a chaotic restaurant plate — a curry, a mixed stir-fry, anything where the ingredients and oils are hidden — is still a weaker guess than a careful manual entry, and the responsible move there is to log by hand. PlateLens is also mobile-only, with no desktop client, and its free tier caps how many photo scans you get per day, though manual logging stays unlimited. For packaged foods and home cooking, where most people do most of their eating, those caveats rarely bite. For the long tail of opaque restaurant meals, they do.
It’s also worth keeping the accuracy conversation in proportion. Federal labeling rules already allow a packaged food’s stated calories to be off by a meaningful margin, so chasing perfect precision past a point is a fool’s errand. Consistency you’ll sustain beats precision you’ll abandon — which is precisely the axis on which photo logging has pulled ahead.
The bottom line
The useful question in 2026 is no longer “do AI calorie trackers work?” They do. It’s “which one removes the friction that makes people quit without quietly wrecking the accuracy?” On that question, the evidence and the hands-on experience point the same direction: for most people, photo-first logging — and PlateLens specifically — has crossed from gimmick to the most defensible default, and in doing so has done something the category spent a decade failing to do. It has made the gold-standard advice — log everything, accurately, forever — into something a normal person will actually keep doing.