AI Fails Where It Matters Most: Food, Minds, Chemistry

AI Can Name Your Food But Can't Count Its Calories — and That's Dangerous

Your phone can name a croissant. Ask it how many calories are in it, and suddenly it's guessing — loudly and confidently.

A team built a benchmark called OmniFood-Bench, drawing from a dataset of 100,000 food images, to run six major AI vision models — including GPT-5.1 and Google's Gemini — through three levels of food-related questions: name the dish, estimate the portion size, then advise someone with a health condition on whether to eat it. Level one? The models are nearly human-level. They know their food names. Level two is where things fall apart. Ask them how much something weighs — which is what you actually need for a calorie count — and accuracy collapses. Think of it like a chef who has memorized every dish on every menu but has never once used a scale. They'll give you a confident-sounding answer regardless. Level three is where it gets genuinely worrying. When the models were told 'this person has type-2 diabetes' and shown a photo of a risky food, several of them hallucinated — meaning they invented — safe-sounding dietary advice. Correct ingredient identification, wrong health conclusion. The researchers call this a 'semantic-physical gap': knowing the name of something is completely different from understanding its nutritional reality. The gap exists because visual appearance simply does not reliably predict what's actually in food. Why does this matter? Diet apps and AI health assistants are being built on exactly this kind of model. If the underlying system is overconfident about what it doesn't know, someone following its advice could be actively harmed. The catch: this paper describes the problem, not a solution. None of the six models is declared safe enough. The benchmark also doesn't yet cover video or varying portion depths, and establishing reliable ground-truth calorie counts for real-world food remains genuinely hard.

AI Still Trails Real Psychiatrists by 37 Points — a New Test Shows Exactly Where

Put the best AI through a virtual psychiatric ward and it trails a real doctor by the same gap a driving-test student trails a licensed instructor.

A team working across six medical centers built a virtual psychiatric hospital from 1,193 real, de-identified patient records covering 76 different mental health diagnoses. They then ran several AI models through the complete clinical workflow a trainee doctor follows: interview the patient, examine their mental state, form a diagnosis, plan a treatment. The whole thing, not just the trivia-quiz version of medicine. Think of it like a driving test. You can know every road rule in the book. That is not the same as handling a nervous passenger, an unexpected cyclist, and a confusing junction simultaneously. The AI models, it turns out, know the road rules very well. The junction is harder. The numbers: the strongest AI model trails real clinicians by 37 percentage points on objective measures — accurate diagnosis, appropriate treatment planning. It also trails medical trainees by 27 points. The specific bottleneck the team identified is what clinicians call a mental status examination: reading how a patient is actually behaving in the room — their affect, their coherence, their silences — rather than what they say in words. Text-based signals, the AI handles reasonably. The full human encounter, not yet. The team also built a new evaluation tool called MentalEval — five AI judges trained on expert clinical feedback — which agreed with human expert ratings at a correlation of 0.94. That matters: it means future tests on newer models will be reliable and repeatable. The catch: these are simulation cases drawn from records, not live patients. How performance maps onto a real ward, with real complexity and real stakes, is still unknown. The 37-point gap is a starting measurement, not a ceiling.

Making AI Argue With Itself Cuts Chemistry Hallucinations by 79%

What if every time an AI was about to answer a chemistry question, it had to defend that answer to a panel of skeptical colleagues first?

A team built what they call G-Frame: a multi-agent training system inspired by game theory, where several AI 'players' interact to generate chemistry knowledge. The analogy that fits best: imagine a kitchen where every cook must justify each ingredient substitution to a skeptical sous-chef before the dish leaves the pass. The sous-chef can reject the answer and send it back for revision. Nothing exits without surviving scrutiny. Using this approach, the team synthesized 363,000 worked-through chemistry examples — full chains of reasoning, not just question-answer pairs — and used them to train a relatively small model called OmniChem, with 7 billion parameters. For reference, that is modest: systems like GPT-4 are estimated to be vastly larger. Size is not what they were optimizing for. The result: OmniChem hallucinates 79.46% less often than the base model it was built on, as measured by a chemistry-specific evaluation tool the team also developed. On public chemistry benchmarks, it matches GPT-4o mini. Why does this matter beyond chemistry? Chemistry is one of the hardest domains for AI errors because mistakes compound — wrong reagents, wrong safety warnings, wrong reaction conditions. A smaller, more careful model that outperforms a larger one on reliability is relevant to any lab or researcher who cannot afford large commercial API costs. The catch: two of the three evaluations used benchmarks the same team built, which creates obvious potential for optimism. The 79% reduction is relative to the base model, not to a human expert. The absolute error rate is not zero. Independent replication on held-out test sets by a different group is the next thing to wait for.

Three papers, one pattern: AI that is confident where it should be uncertain, and wrong in ways that could hurt someone. The food model knows dish names but invents calorie counts. The psychiatric AI recites clinical knowledge but misreads the patient sitting in front of it. That is not random noise — it is a structural tendency. These models are very good at surface pattern-matching and genuinely shaky at grounded, contextual, physical reasoning. The G-Frame chemistry result points at one avenue: force the model to justify itself before committing to an answer. That is an old idea in human institutions — peer review, surgical checklists, the sous-chef at the pass — and it appears to transfer to AI training. But a 79% reduction in one narrow chemical domain is not a general fix for confident wrongness across food, psychiatry, and everything else. What these three papers collectively give us is better measurement. We now know the gap is 37 points in psychiatry. We know exactly which tier of food reasoning breaks. That precision is genuinely useful. It just does not close the gap on its own.

The OmniFood benchmark is now public, so expect competing teams to test newer models against it over the coming months — watch for whether any model clears the safety-critical advisory tier reliably. For MentalHospital, the key open question is whether the 37-point gap shrinks when models are fine-tuned on the clinical encounter format rather than tested cold. And for G-Frame: the test I would want to see is OmniChem evaluated by a chemistry team with no connection to the original paper.

• Wikipedia: Hallucination in artificial intelligence
• OmniFood-Bench paper on arXiv
• MentalHospital paper on arXiv