Fix the input, not just the number
How to correct an AI calorie estimate
When a photo-based result looks wrong, the best correction explains what the image could not reveal.
Published · Updated
Short answer
Correct an AI calorie estimate by stating the specific mistake: food identity, portion size, hidden cooking fat, sauce or topping, or the fraction actually eaten. A concrete correction such as “that was 200 grams and I ate half” is more useful than saying “the calories are wrong.”
Use the smallest specific correction
You do not need to describe the whole meal again. Correct the uncertain part and preserve what was already understood.
| Problem | Weak correction | Useful correction |
|---|---|---|
| Wrong food | “That's wrong” | “The white portion was tofu, not chicken” |
| Wrong quantity | “Too high” | “I ate half of the rice shown” |
| Hidden fat | “More calories” | “It was cooked with about one tablespoon of olive oil” |
| Missing topping | “You missed something” | “Add two tablespoons of peanut sauce” |
| Leftovers | “I didn't finish it” | “I ate roughly two thirds of the plate” |
Correct ingredients before adjusting the total
Changing only the calorie number can make the macro breakdown internally inconsistent. When possible, identify the ingredient or portion that caused the difference.
For a packaged item, use the label and amount consumed. For a homemade recipe, the ingredient weights are stronger evidence than visual appearance.
When a range is more honest than a point estimate
Restaurant dishes often contain unknown oil, butter, sugar, and sauces. If those cannot be recovered, a reasonable range can communicate uncertainty better than a precise-looking total.
The purpose of a correction is to make the log more useful—not to create certainty the available evidence cannot support.
Do not repeatedly tune an estimate until it matches the number you hoped to see. Use new evidence: a label, a measured portion, a recipe, or the amount left uneaten.
How correction works in Keola
Keola is designed for conversational correction. After a meal is logged, reply with the replacement food, changed amount, or missing detail. The corrected meal can then update the day's totals.
This interaction is one reason WhatsApp is useful: the original photo, estimate, and correction remain in the same conversational context.
Frequently asked questions
Should I correct every small difference?
Not necessarily. The appropriate precision depends on your goal. Correct details that materially change the usefulness of the log.
What if I know the nutrition label?
Use the label and quantity consumed. That is stronger evidence than a visual estimate.
Can a tracker infer cooking oil from a photo?
Sometimes it may infer that oil was likely used, but the amount is usually uncertain unless you provide it.
Should I use a calorie range?
A range can be more honest when important ingredients or quantities remain unknown.
Test the interaction, not the claim
Send one real meal in WhatsApp. No account or payment card is required for the first analysis.
Try one meal in WhatsApp