Why AI confidently states incorrect figures

The unpleasant feature of machine responses is not that they can be incorrect—reference books and labels can be wrong too. It is that correct and incorrect answers look identical: the same number of digits, the same tone, the same speed. Understanding where this consistency comes from and what to do about it is more useful than arguing over accuracy percentages.

The model always provides an answer

Language or vision models are designed to provide the most probable continuation—for an image of a dish, this is a plausible estimate. There is no "I don't know" state in this scheme: even in a photo where neither the dish nor the plate size can be discerned, a weight and calorie count will be provided. This is not because the model is confident, but because it has no way to formulate a refusal.

This leads to a practical consequence. Silence or a disclaimer is a signal you should not expect. You have to verify the plausibility of the answer from the outside, and below we break down exactly how to do that.

The same plate, two different answers

A repeat request for the same photo often yields a different number. The reason is not a malfunction: generation is probabilistic, and a small variance is built into the procedure itself. It is useful to use the magnitude of this variance as a free check—if two answers differ by half, it means there is truly nothing to measure in the photo.

The same technique is used in research. When testing chatbots on sports nutrition questions, the reproducibility of the answer when repeating the same question was evaluated separately, and it turned out to be one of the weak points: the authors describe the accuracy as moderate and inconsistent between models.

Systematic versus random error

This distinction is more important than the overall accuracy figure. A random error fluctuates around the true value and is partially canceled out over the long run. A systematic error is biased in one direction and never cancels out—no matter how many days you count.

In a test of an image recognition app against a weighed food diary, the bias turned out to be systematic: energy was overestimated by an average of 154 kcal per day, fat by 10.4 g, carbohydrates by 33.8 g, while fiber was underestimated by 4.5 g. Thirty-six people, ten days in a row, both records simultaneously.

Reviews of image-based methods describe another pattern: the bias grows along with the portion size. A small plate is evaluated reasonably well, a large one noticeably worse, and almost always in the direction of underestimation.

What follows from this. If you use the same tool and compare your days against each other, the systematic bias is reduced: it is present both yesterday and today. If you compare a figure from an app with a recommendation from an article or someone else's result, the bias remains in full.

Five signs of an answer that should be double-checked

• A round number. "500 kcal" for a photo of a home-cooked plate is almost always a rounded guess, not a calculation.
• An overly precise number. "487 kcal" creates the impression of a measurement where there was none. Three significant digits with a margin of error of one-third is decoration, not information.
• A dish is named that cannot be identified in the photo. If you cannot see for yourself in the photo whether it is chicken or pork, the model cannot see it either—but it will name it.
• Oil is not mentioned. Frying adds 100–200 kcal per serving and is not visible. An answer where a fried dish is valued the same as a boiled one is underestimated.
• The answer did not change after you added important details. If adding "dressed with two tablespoons of mayonnaise" changed nothing, the note was not taken into account.

What to do with this in practice

Three techniques cover most cases, and all three are free.

Compare with an anchor. Keep a few reference values in mind: a tablespoon of oil is about 110 kcal, a slice of bread is about 80, a 150 g serving of boiled buckwheat is about 200. An answer that differs from the anchor by several times is pointless to discuss; it needs to be clarified.

Add the invisible. Cooking method, dressing, fat content of dairy, sugar in a drink. These are exactly the things that are not reflected in the photo but are decisive for calorie content.

Look at the week, not the meal. The error of an individual lunch is large; the weekly average is more stable because the random component is partially canceled out. It makes sense to compare the weekly average with the daily calorie requirement, rather than an individual dinner.

And a general rule that is older than all neural networks: an estimation tool does not replace a scale where a scale is needed. If the task is an accurate deficit at a low body weight, where 200 kcal per day matters, you should calculate based on weighed home recipes, and the photo remains for everything else.

Why "fine-tuning the model" does not resolve the issue

Some errors do indeed go away with model improvements: dish recognition, rare categories, cuisines that were not in the training set. But the main component is not in the model. A flat image does not contain information about what is hidden under the top layer, how much oil was absorbed, or what the fat content of the ground meat was. This is a limitation of the image, not the algorithm, and it is not resolved by data volume or network size.

Therefore, an honest formulation for any tool of this kind sounds the same: it provides an estimate, not a measurement. An estimate is useful exactly where a decision is made at the "a lot or a little" level, and useless where a decision is made at the "exactly how much" level.

Count it from a photo

Frequently asked questions

Why does the same photo yield different answers?

Response generation is probabilistic, so a small variance is inherent to the process itself. This can be used as a check: if two answers for the same image differ significantly, it means there is truly insufficient data in the photo for an assessment.

Can the model say it doesn't know?

As a rule, no. It provides the most plausible continuation, and a refusal is not built into this scheme: even for an unreadable image, a number will be provided. The absence of a disclaimer does not imply certainty — it simply cannot be expressed.

What is a systematic error and why is it worse than a random one?

A random error fluctuates around the true value and is partially canceled out over the long term, while a systematic error is biased in one direction and is never canceled out. In tests of image recognition applications, the energy bias was around 150 kilocalories per day and increased with portion size.

How can I quickly tell if the answer is underestimated?

Check if invisible items were accounted for: oil used for frying, dressing, sugar in a drink, or the fat content of dairy products. If a fried dish is estimated to be roughly the same as a boiled one, the answer is underestimated by 100–200 kcal just from the oil.

Is it worth using such an estimate at all?

It is useful where a decision is made at the level of "a lot or a little": to see that a lunch was not four hundred kilocalories, but over a thousand. For precise deficit calculations, where tens of calories matter, you need a scale and a recipe.

Read next

• How a neural network recognizes food in a photo — A photo of a plate turns into a number on the screen in a few seconds, and from the outside, it looks like a measurement.
• AI nutritionist and nutrition chatbots: what can you trust — Asking a chatbot about your diet has become easier than making an appointment with a specialist, and the answer comes detailed, polite, and structured.
• Calorie counting apps: how they work and where they go wrong — A calorie counter looks like a calculator, and people expect calculator-like precision from it.

This article is for general information. It is not medical advice, a diagnosis, or a prescription for treatment or a diet, and it does not replace a consultation with your doctor. If you have a health condition, are pregnant, take medication, or follow a diet prescribed to you, decisions about food belong with your doctor.

Figures from regulations, guidelines and studies are given as they stood when this article was prepared and may since have changed; check them against the primary sources. This article is not advertising, an offer, or individual advice, and neither the author nor the site owner is responsible for decisions taken on the basis of it.