A new benchmark exposes a wide gap between what AI vision models see on a plate and what they actually understand about it.
Researchers introduced OmniFood-Bench, built from a dataset of 100,000 food images, to test whether large vision-language models can do more than identify dishes. The benchmark runs models through three tiers: recognizing ingredients and cooking methods, estimating portion sizes and nutrients, and generating advice for users with specific health conditions. The results are blunt — models score near-human on naming food but collapse when asked to estimate mass or tailor advice for high-risk profiles like diabetics. The researchers call this the "Semantic-Physical Gap," a mismatch between labeling what something is and reasoning about what it contains.
The gap matters because AI-assisted dietary tools are already reaching consumers, and the failure modes here are not benign. Hallucinating a safe recommendation for a diabetic user is not a classification error — it is a patient safety problem. Benchmarks that only test food recognition have been obscuring this risk.
Food AI has long oversold its clinical readiness. The pattern of "impressive demo, brittle edge case" is familiar from medical imaging and radiology tools that performed well on curated test sets and poorly in deployment. OmniFood-Bench at least gives the field a harder target to aim at — though a benchmark only changes outcomes if model developers actually train against it.