Can AI count calories from a photo of my food?

It can estimate them. A model reading a photograph can identify the foods and approximate portions, which is usually accurate enough for tracking trends over weeks. It cannot measure — it cannot see oil absorbed during cooking or know a sauce's exact composition — so a good implementation reports a confidence level and shows the components it derived the estimate from rather than presenting a guess as a measurement.

How accurate is AI food logging?

Accurate enough for trends, not for precision goals. Estimates from a description or photo are typically close on calories and macros for simple, recognisable meals, and less reliable for mixed dishes, restaurant food, and anything where cooking fat is hidden. Adding detail to the description — portion sizes, cooking method — improves it more than anything else you can do.

Why do some nutrients show as unknown instead of zero?

Because a zero is a claim that the food contains none of that nutrient, and an estimate from a photo rarely supports that claim. Recording unknown values as zero would quietly average into weekly totals as though measured, producing apparent deficiencies that do not exist. Leaving them unknown, skipping them in totals, and reporting how many entries knew each nutrient is the honest handling.

Do I need a separate app to track calories this way?

No. That is the point of logging through a messaging thread — the food log lives in the same conversation you already use, so there is no app to open and no database to search. Entries are still viewable and editable in a dashboard when you want to review or correct them.

Guide

Log your food by texting it

Food logging fails on friction, not on maths. Sending a photo or a sentence to a thread you already have open removes the step where most people quit — provided the estimate tells you honestly how confident it is and what it was derived from.

By Alec LeyfermanAugust 21, 20267 min read

Almost nobody quits tracking their food because the arithmetic got hard. They quit because logging a meal means unlocking a phone, finding an app, searching a database, picking the closest of nine entries for grilled chicken, and guessing at a serving size — five times a day, forever.

The interesting question is not how to make that flow prettier. It is whether the flow needs to exist at all.

Three ways in, one path through

If your assistant lives in a message thread, logging a meal can be a message. In Cael that means three routes to the same place:

Text it. "chicken burrito bowl, brown rice, double chicken, no cheese" gets estimated and logged.
Say it. A voice note is transcribed and handled exactly like the typed version — nothing extra to set up.
Photograph it. Send a picture of the plate, with or without a caption, and the model reads the food off the image.

What matters technically is that these are not three features. A photographed plate is just a turn where the model happens to be looking at a picture before it calls the same logging tool the typed version calls. One estimator, one schema, one confirmation. Separate paths drift apart within a month, and then the photo route quietly starts producing different numbers than the typed one.

The whole label, not just a calorie count

Calories are the number people ask for and rarely the number that changes anything. A log worth keeping records what a nutrition label records: protein, carbohydrate and fat, then the parts a label breaks out — saturated and trans fat, fibre, sugar and added sugar — and then the micronutrients that actually vary day to day, including sodium, cholesterol, potassium, calcium, iron, and vitamins A, C, and D.

That distinction is also why food does not fit in a generic tracker. A weight or a mood is one number per entry. A plate is twenty numbers that only mean anything together, plus the description they came from.

What an estimate can honestly know

This is where most AI food logging goes wrong, and it is worth being blunt about. A photograph of a curry does not contain enough information to determine its vitamin D content. Nothing does, short of a lab.

There are two ways to handle that. One is to output a number anyway, because a filled field looks more finished than an empty one. The other is to leave it unknown and say so.

The second rule that follows from the first: say where a number came from. Every entry keeps the components it was derived from — "6oz chicken, 1 cup rice, olive oil" — along with how it was captured and an honest confidence rating. A photo estimate is not a label read, and it should not be presented as one.

The point of showing all that is correction. An estimate you can inspect is one you can fix when it is wrong. An estimate you can only distrust is worse than no estimate at all.

Where this approach loses

Packaged food with a barcode. If the item has a label and a scanner can read it, scanning beats estimating every time. That is a measurement, not a guess.
Precision goals. Anyone weighing food on a scale for a competition or a medical reason needs measurements, and should not be working from photographs.
Ambiguous plates. A photo cannot see the oil in the pan or the butter under the steak. Adding "cooked in about a tablespoon of oil" to the caption moves the estimate more than any model improvement will.
Long-run device integration. Dedicated fitness platforms sync with scales, watches, and rings. A messaging agent does not.

Where it wins

Against the realistic alternative, which is not a perfect log but no log at all. A rough estimate captured in four seconds from a thread you already had open beats a precise entry you were going to skip. Consistency over months is what makes a food log useful, and friction is what ends it.

If you want the comparison against the dedicated tools, Cael vs MyFitnessPal covers where each one genuinely wins.

Frequently asked questions

Can AI count calories from a photo of my food?: It can estimate them. A model reading a photograph can identify the foods and approximate portions, which is usually accurate enough for tracking trends over weeks. It cannot measure — it cannot see oil absorbed during cooking or know a sauce's exact composition — so a good implementation reports a confidence level and shows the components it derived the estimate from rather than presenting a guess as a measurement.
How accurate is AI food logging?: Accurate enough for trends, not for precision goals. Estimates from a description or photo are typically close on calories and macros for simple, recognisable meals, and less reliable for mixed dishes, restaurant food, and anything where cooking fat is hidden. Adding detail to the description — portion sizes, cooking method — improves it more than anything else you can do.
Why do some nutrients show as unknown instead of zero?: Because a zero is a claim that the food contains none of that nutrient, and an estimate from a photo rarely supports that claim. Recording unknown values as zero would quietly average into weekly totals as though measured, producing apparent deficiencies that do not exist. Leaving them unknown, skipping them in totals, and reporting how many entries knew each nutrient is the honest handling.
Do I need a separate app to track calories this way?: No. That is the point of logging through a messaging thread — the food log lives in the same conversation you already use, so there is no app to open and no database to search. Entries are still viewable and editable in a dashboard when you want to review or correct them.

Hand Cael your first task tonight.

Cael runs in iMessage and Telegram. Text it something small and reversible — a table to find, a price to watch — and see what comes back.

Start 7-day trial

$0 today · Card required · Cancel anytime