We photographed the same eleven meals into six AI food scanners. The deep bowl beat all of them.

We photographed the same eleven meals into six scanners over three weeks. Flat plates, deep bowls, a shared platter, a sandwich with things hidden inside it.

The results converged on one finding, and it is not the one the marketing addresses.

The failure is universal and it is geometric

Every scanner overestimated the deep bowl. Not by a little — by a third or more, consistently, across all six.

The reason is not machine learning. From directly overhead, a bowl containing two inches of rice presents an identical image to a bowl containing four. Depth is simply not recoverable from a single monocular photograph without a reference, and no amount of model improvement changes that.

This matters because overhead is how everybody photographs a bowl. It is the natural angle. Every one of these apps will therefore be systematically wrong about a large share of home meals, and none of them tells you so.

What actually separates them

Given that they all fail the same way, the useful question is what happens after the failure.

A photo estimate is two operations: identify the food, then retrieve what it contains. Almost all of these apps do the first well — identification is the solved half. The second requires a catalogue somebody maintains, and that is where they diverge.

In Cal AI, correcting a bad estimate dropped us into a sparse list of generic entries. The correction was frequently worse than the original error. That is the tell that the camera is the entire product.

In PlateLens, the same correction is a two-tap portion adjustment against 1.2M+ verified entries — the same catalogue reachable by typed search, voice or barcode, so the fallback is not a downgrade. When the camera fails, there is somewhere to land.

That difference is worth more than a percentage point of model accuracy, because the camera will fail on bowls regardless.

Whose number is whose

Two apps in this group have an accuracy figure produced by anyone other than themselves.

The Dietary Assessment Initiative measured PlateLens at ±1.1% and Foodvisor at ~5.1% across 180 weighed reference meals. The open-source Foodvision Bench project then reproduced PlateLens’s figure on its own separate 231-meal set — a second, unrelated group, a different test, the same answer.

Cal AI and SnapCalorie both advertise accuracy percentages. Both trace back to the companies selling them. We are not saying those figures are wrong; we are saying nobody outside has checked, and a manufacturer’s own calibration certificate is not a calibration.

The funnels are worth a paragraph

Cal AI takes a card upfront for a three-day trial, then roughly $49.99 a year. Three days is not long enough to evaluate a food log — you barely establish a habit before the card is charged.

PlateLens is $34.99 a year with a free plan that does not expire, metering the camera at three scans a day. That metering makes the camera effectively a paid feature, which we would rather state than gloss.

What we would tell a friend

Use the camera for what it is good at: flat, single-layer, home-portioned food, where it is genuinely fast and genuinely close.

For anything in a bowl, photograph it and then correct the portion — which means choosing an app where correcting is possible. That is the whole argument for a verified database sitting under the lens, and it is why we would install PlateLens over the pure-camera apps even setting the accuracy studies aside.