Breaking01 / AI
AI calorie estimates fail to scale with stated food weights
Multiplying each model’s calorie estimate per unit of weight by the stated food weight reduced errors for small and large portions.

Researchers published an evaluation of calorie estimates from seven AI models on arXiv on 20 Nov 2026. They compared answers across stated food weights and excluded models whose weights were private.
The researchers tested language models with publicly available weights. They asked the models to estimate calories for 168 foods at different stated weights.
The abstract reported that none of the models’ calorie estimates changed in proportion to food weight. Four models overestimated calories for small portions and underestimated them for large portions.
The researchers multiplied each model’s estimate of calories per unit of weight by the stated food weight. This reduced errors for both small and large portions.
The researchers recommended the calculation for food-logging apps. The abstract did not report results from using it in an app.