The short answer

No, not in the way people mean when they ask.

AI calorie counters produce estimates. So do food databases. So does the nutrition label on a packet, which is legally permitted a meaningful margin of error. Nobody involved in this is measuring anything — everyone is estimating, and the honest question is which estimate is built on better information.

What "accurate" would even mean

To know the true calorie content of a meal you would need to know the exact mass of every ingredient, the exact fat absorbed during cooking, and the actual composition of the specific produce used. Restaurants do not know this. Your family does not know this. You do not know this about the dal you made on Sunday.

So the real comparison is not estimate versus truth. It is estimate versus estimate.

  • A packaged food database entry is very good, because someone measured a standardised product.
  • A user-submitted database entry for a home-cooked dish is one stranger's guess about their kitchen, presented to you as a fact.
  • A photo estimate is a model's guess about what is on a plate and how much of it there is.
  • A described-meal estimate is a model's guess, informed by whatever you chose to tell it.

Only the first of these is meaningfully "accurate", and it only covers food that came in a packet.

Where photo-based estimation struggles

Photo apps have improved a lot and are genuinely decent at recognising common, plated, visually distinct food. Two problems remain, and they are structural rather than temporary.

Volume is hard to see. Depth, plate size and camera angle all distort apparent quantity. A small katori and a large one photographed from above look very similar. Since calories scale directly with quantity, an error here is an error in the final number, one-to-one.

Fat is invisible. This is the bigger one. A dal cooked with a teaspoon of oil and the same dal with three tablespoons of ghee are visually near-identical and differ by 200 calories a serving. No amount of model improvement fixes this, because the information is simply not present in the image.

Both of these are things you know. You know it was two rotis. You know your mother cooks with more ghee than the restaurant.

Where described-meal estimation struggles

To be even-handed, it has its own failure modes.

You have to describe it reasonably. "Some curry" gets you a generic answer. "A bowl of chicken curry, home-made, medium oil" gets you a much better one.

Unfamiliar restaurant dishes. If you cannot name it or describe its ingredients, a photo genuinely asks less of you.

Portion honesty. The model believes you about quantity. If you routinely say "one bowl" for what is closer to two, the estimate is wrong and it is not the model's fault.

The question people should be asking

Here is the thing that gets lost. In practice, almost nobody fails at calorie tracking because their numbers were 12% off.

They fail because they stopped logging.

The typical arc is familiar: four days of diligent database searching, then a busy day, then a skipped meal, then a skipped day, then the app sits unopened. Three weeks of perfectly precise data followed by nothing is worth considerably less than six months of decent estimates.

This reframes the whole accuracy debate. A system that is 90% accurate and used daily for a year beats a system that is 97% accurate and abandoned in a fortnight. It is not close.

Consistency beats accuracy

There is a second reason to relax about the exact numbers.

If your estimate for a roti is consistently 15% low, but it is consistently 15% low, your day-to-day comparisons still work. The trend still tells you whether you are in a deficit. Your weight over three weeks corrects the bias for you — you simply eat a bit less than the target you set, and the number on the scale tells you so.

What breaks this is inconsistency: counting the ghee some days and not others, logging lunch carefully and estimating dinner from memory a day later. That produces noise you cannot correct for, because you cannot see it.

Consistent estimates are self-correcting. Sporadic precise ones are not.

What to do about it

Log everything, roughly. A rough log of the whole day beats a precise log of two thirds of it.

Be consistent with your assumptions. Pick a number for your usual roti and stay with it.

Correct against the scale, not the app. After two to three weeks, your weight trend tells you whether your intake estimate is systematically off. Adjust the target, not the individual entries.

Track the fats. Across almost every dish, added fat is where the largest correctable error lives.

How BetterCal approaches it

BetterCal says plainly in the app that these are estimates. It does not present a decimal place it has not earned.

What it does instead is make the estimate as informed as possible — using your country's actual cooking norms, ingredients, fats and portion sizes rather than a single global average — and make logging fast enough that you keep doing it. Corrections are remembered, so the estimates converge on your kitchen over time.

The bet is that a good estimate you record every day for a year is worth vastly more than a perfect number you stop entering in March. We are fairly confident that bet is right.

Related: why photos are the wrong input · what to look for in an AI calorie tracker · why people quit calorie tracking