Estimating calorie intake from a photo of a meal is an appealing shortcut for dietary tracking. General-purpose multimodal large language models make such estimates easy to obtain, but how accurate are they? A new preprint from a feeding trial suggests the answer depends on whether you look at the group average or the individual case.
The paper, titled "Small Average Error, Wide Individual Error," reports that when general-purpose LLMs estimated energy intake from meal photos, the average error across meals was small. However, that average masked a wide spread: individual estimates could be off by a large margin.
These findings matter for real-world dietary assessment, where an individual's error is more relevant than a group average. A tool that is accurate on average but unreliable for a given meal could mislead users who depend on it for daily decisions.