AI and the World's Hard Problems · Medicine & drug discovery · arXiv · 2026-10-20
Small Average Error, Wide Individual Error: General-Purpose LLMs Estimating Energy Intake from Meal Photos in a Feeding Trial
Photo-based energy estimates could ease dietary assessment, but most evaluations compared estimates with energy served rather than energy eaten. The authors tested six general-purpose multimodal LLMs on 802 meals from a controlled feeding trial with weighed intake, comparing before-only photos, before-and-after subtraction, and joint input. Subtraction under-estimated intake for five of six models, and the authors conclude that a low average bias is not evidence of accurate individual estimates, which should support rather than replace assessment by a dietitian.