← All food signals
Trust signal

Machine Dietetics

Chatbot diet advice and photo-to-nutrient AI are now tested against dietitians and weighed meals. Errors cluster on B12, mineral doses, high-risk life stages, fat and salt — the costliest mistakes.

What is it?

Researchers now score chatbot diet answers and meal-photo apps against dietitians and weighed plates. The models are mostly safe on everyday questions and wrong in the same places: B12, mineral doses, fat, salt, specialist menus. Whether those weak points shrink with newer models stays open.

Why now?

Trust in food advice has moved from people to machines faster than anyone checked it. Peer-reviewed tests now name the model version and the nutrient it misses. The outside check has arrived before any rulebook has. Each study so far finds errors gathering on the same few high-stakes points.

Who’s building it?

Nutrition researchers score chatbots and photo apps; model makers and app developers ship the answers being scored.

Nutrition researchersRegistered dietitiansModel makersMeal-photo appsJournal reviewers
The full read
Why now, who is building it, where the market goes, and what stays unsettled.
Get weekly →
01 · The market
Who’s in it
Builders — model makers and meal-photo apps answering food questions.
Buyers — people asking a chatbot what to eat, or photographing a plate for numbers.
Gatekeepers — dietitians and peer reviewers who score the answers against evidence.
What’s at stake
Whether a fluent answer can be trusted on the B12, dosing and life-stage questions where a wrong one does harm.
Fat and salt estimates that miss by nearly half.
Specialist menus for children that break the prescription.
What’s unsettled
Whether newer model versions close the same gaps older ones showed.
Whether real phone photos perform worse than controlled lab shots.
Whether any dietetics body or regulator sets a public accuracy bar for diet answers.
02 · How the demand matureda progression, not a timeline
Ask
A person puts a diet or food-safety question to a chatbot.
Snap
A meal photo becomes a calorie count and a macro breakdown.
Score
Researchers test those answers against dietitians, weighed meals and specialist guidance, nutrient by nutrient.
Rely
Whether people learn where the weak points sit decides what trust survives.
03 · Where the market goesthe forks, not a prediction
→Close the gap or keep it: newer models may fix B12, dosing and fat errors, or the same weak points may persist.
→Dietitians may become the standing human check on machine advice for children and other high-risk diets.
→Meal-photo apps may publish their error nutrient by nutrient, showing buyers where calorie counts are weakest.
→Advice may split into a general layer and a specialist one.
Machine diet advice is already mostly right; the open part is whether anyone, maker or user, ever sees which answers fall in the risky few.
04 · The Verification Stack deskadjacent markets, moving now
The Skin Metric
Another AI grader, checked from outside.
AI skin grades matched dermatologists exactly 51.4% of the time.
Skin Research and Technology via PubMed Central
Metabolism, Made Visible
Body data reaches consumers without a prescription.
Dexcom’s Stelo OTC glucose monitor cleared for children.
Fierce Biotech
The Verification Reflex
Machine output meets a public check, elsewhere.
UK Electoral Commission piloted deepfake detection before May 2026 elections.
Tech Policy Press
05 · The Wire
01
Five chatbots answered 58 questions; most were rated safe.
Frontiers in Public Health ↗
02
Three chatbots missed prescribed diet targets for children with epilepsy.
Journal of Human Nutrition and Dietetics ↗
03
On real-world meal photos, Claude Sonnet 4.6 missed energy by a median 33.5%, fat by 46.9% and salt by 55.8%.
Nutrients ↗
04
Against 15 weighed hospital meals, top AI models matched dietitians on energy and carbohydrate but all AI models significantly overestimated fat.
Nutrients ↗
05
The best AI mushroom identifier failed in almost 15% of cases on 100+ real-world photos of nearly 60 species.
npj Science of Food ↗
The panel on this signal
They agreed on
–Most chatbot diet answers tested were rated safe, and the unsafe ones gathered on a few predictable topics.
–Photo-based AI misses fat by more than any calorie count.
–Children’s therapeutic diets are where tested chatbots broke the prescription outright.
The food-composition scientist disagreed on
–Fat errors are partly a reference problem: hidden cooking oil defeats a photo, and composition tables vary too.
Know more about the method →
✦How this signal was built
01 · Panel
A Delphi panel of 4 AI analysts, each reasoning independently.
→
02 · Debate
They challenge and revise each other’s reads over two rounds.
→
03 · Evidence
Every claim is tied to a cited, linked source on The Wire.
→
04 · Synthesis
A moderator writes it into one read, disagreement and all.