The honest framing
Photo and voice logging produce estimates. Identifying what's on the plate works remarkably well; estimating how much is genuinely hard — a camera can't weigh rice. Expect the identity to be right and the portions to deserve a glance.
That's exactly why the review screen exists: everything is editable before it enters your diary, and the foods that matter most to you take two seconds to correct.
The three badges, and what each one means
A detected food arrives in one of three states, and they are not the same problem:
| Badge | What it means | What to do |
|---|---|---|
| Needs match | We know roughly what it is but couldn't tie it to a catalog food | Tap it and pick a more specific food |
| Review | We're not confident what it is. Tapping opens search preloaded with the alternatives we considered | Confirm or choose one |
| AI estimated | Identified, but the numbers come from the model rather than a catalog entry | Trust it less than a barcode; edit if it's a staple |
There's also a guard you'll never see named, which explains most surprising "Needs match" results: before accepting a catalog food, the app compares its energy density — calories per gram — against what the analysis estimated. If they diverge by more than 3×, the match is rejected and the food goes to review instead. It's the check that stops "red pepper flakes" from being logged as raw bell pepper: the words match, the density doesn't, and density doesn't care about portion size.
Making estimates better
- Say more. In voice (and photo descriptions), explicit amounts are respected — "200 grams of rice" beats "some rice."
- Shoot clearly. One plate, decent light, shot from above-ish. Buried or blended ingredients are the hardest.
- Prefer precision where it's cheap. Packaged products deserve the barcode or label scan — label data is exact in a way estimation never will be.
AI analysis runs against an hourly allowance — ten photo analyses and ten voice logs an hour — which no ordinary day of logging comes near, but a long session of retries can. If you hit it, the app tells you how many minutes remain.
Does imperfection break the math?
Less than you'd fear. Your expenditure is calibrated against your weight trend — consistent logging, even imperfect, still produces a useful estimate. Consistency beats precision; both beat neither.
The one hard line
Never rely on AI logging for allergens or anything else where a wrong answer could harm you. That information belongs to the product label, read with your own eyes — see our Health Disclaimer.