Close, most of the time, and less so on cooked foods. In Kove's own test of 37 foods written with their weight, an AI model guessing on its own was within 10% of the reference figure 73% of the time. Kove found 99% of them in a food table, and every figure it logged was within 10%.
The misses are the foods that change most with cooking: cooked rice was off by 16%, and legumes, meat, fish and eggs by 12-13% at the median. From a photo alone it is harder for any model: one 2025 study found ChatGPT and Claude about 36% off on energy (Fridolfsson et al., 2025), and another found ChatGPT closer once it was told the ingredients and their amounts (Rodríguez-Jiménez et al., 2025).
Kove's test ran on 31 October 2026, each food 5 times. The method and every food.