How accurate are AI calorie counters? What the research says

Updated: 2026-11-04

AI photo estimates

• A systematic review of 52 studies found AI calorie estimation error from 0.10% to 38.3%, lower for single, simple foods (Shonkoff et al., Annals of Medicine, 2023; DOI).
• On 52 standardised food photos, general chatbots had a mean calorie error of 35.8% (GPT-4o and Claude 3.5 Sonnet) and 64.2% (Gemini 1.5 Pro); all underestimated large portions (Fridolfsson et al., Current Developments in Nutrition, 2025; DOI).
• Testing nutrition apps' photo recognition, energy error ranged from −3% to −47% and up to +44% depending on the app. One app missed the butter on eggs on toast and under-counted by 35% (Li et al., Nutrients, 2024; DOI).

Manual logging is not exact either

• Against doubly labelled water, self-reported intake was too low by 12–20% on 24-hour recalls and 31–38% on food questionnaires (OPEN study, Am J Epidemiol, 2003; DOI).
• In a small study of people who said they could not lose weight, intake was under-reported by 47% (Lichtman et al., NEJM, 1992; DOI).

What this means

The weak spots are the same for AI and for people: mixed dishes, hidden fat and big portions. That is why Foodmetric shows every detected food with its grams and lets you correct it before saving, and why barcodes are better for packaged food. Consistency matters more than perfection: logging most meals with a known error tells you more than logging a few meals precisely. See seven tips for more accurate photo estimates.

Sources

• Shonkoff et al., Annals of Medicine 2023
• Fridolfsson et al., Current Developments in Nutrition 2025
• Li et al., Nutrients 2024
• Subar et al., OPEN study, Am J Epidemiol 2003
• Lichtman et al., NEJM 1992