Nutrition apps have been telling people what to eat for years. Most build the ideal plate from scratch – nutrient targets met and processed food minimized.
It’s the meal you’re supposed to have rather than the one you actually make.
None of it gets much traction. A new study tried a different approach: start with the meals people already eat, and find the smallest possible change that actually moves the needle.
The advice gap
The problem is real. A poor diet ranks among the largest causes of diabetes, heart disease, and other chronic illness, as one sweeping study of nearly 200 countries showed.
Acting on it at dinner is another thing entirely.
Trevor Chan and Ilias Tagkopoulos, computer scientists at the University of California, Davis (UC Davis), suspected the real trouble was the size of the task.
Many diet apps demand a full overhaul, and the overhaul is what people quit.
The team’s goal was modest by design.
Rather than build a perfect plate from scratch, they wanted to start with food people already eat and nudge it toward healthy eating with as few changes as possible.
Learning from meals
To learn what Americans really put on their plates, the team leaned on a long-running federal survey called What We Eat in America.
The survey had logged more than 135,000 meals reported by over 55,000 adults.
Sorting through all of it, the model grouped meals into 34 everyday meal patterns – the cereal-and-milk breakfast, the deli sandwich lunch, and the pizza dinner.
A generative AI program then learned to build fresh meals that fit each one.
The program tackled two tasks at once. It chose foods that naturally fit together, then adjusted portion sizes to bring each meal closer to federal nutrition guidelines while still resembling something a person might actually eat.
Closer to healthy targets
When the team set its invented meals beside the real ones in the same pattern, the made-up versions came out healthier. They closed the gap to federal nutrition targets by about 47 percent.
The gains were not abstract. Fiber, protein and potassium all climbed in the generated meals, and thin spots in the vitamins filled in, while the meals held the same general look and flavor as the originals.
One number went the wrong way. Sodium crept higher in some lunches and dinners – a plain reminder that no single fix tidies every nutrient at once.
A few simple swaps
Earlier tools could build a healthy menu from scratch. None had pinned down the smallest fix to a meal you already eat – a couple of well-chosen food swaps.
The pair tested changes of one, two or three items per meal. The most common moves were simple: adding vegetables or legumes and pulling out the saltiest or most heavily processed pieces.
The results scaled with effort. A single swap nudged a meal’s nutrition up around 5% while trimming its modeled cost by roughly a fifth.
Allow three swaps, and meals came out about 10% healthier for nearly a third less money.
The meals stayed familiar – a swap might trade a fatty side for beans or fold in extra greens.
It is not about reinventing dinner – it’s just a leaner take on the same meal.
Better than chatbot guidance
A fair question hangs over any AI food tool: why not just ask a chatbot?
The team tested exactly that, pitting its purpose-built model against GPT-4o, the strongest general chatbot available at the time.
Its specialized model won where it counted. Protein, fat and carbohydrate balance – it matched federal targets far more consistently.
The chatbot drifted toward plates heavy in fat and short on carbs. That gap fits a wider pattern.
A recent review of chatbots giving diet advice found their guidance uneven and sometimes wrong, suggesting that building nutritional rules directly into the model may give an edge over free-form chat.
Limitations of the study
The study comes with an important limitation: all of the results exist only in a computer model.
No one has prepared these meals, eaten them, or tested whether the suggested swaps are practical to maintain over time.
The data also have built-in biases. Participants reported their own diets, and people are known to underreport less healthy foods while overstating healthier choices.
The estimated cost savings are similarly based on modeled menus rather than actual grocery receipts.
The authors do not oversell it. They see one clear thread running through the numbers.
“Healthier eating does not have to mean giving up the meals people already enjoy,” the researchers noted.
What could change
What is new here is the size of the lever. A meal does not need a redesign.
One or two well-chosen swaps can move it toward the guidelines and cut the bill at the same time.
The practical upshot is not hard to see. A grocery app could suggest a single swap at checkout instead of a whole new diet, and a public-health program could offer cheaper, healthier takes on the meals people already cook.
Ultimately, the same engine could feed the tools dietitians use, proposing changes a patient might keep up.
The larger lesson lands cleanly: better eating may be less about willpower than about the right small change.
The study is published in the journal PLOS Digital Health.
—–
Like what you read? Subscribe to our newsletter for engaging articles, exclusive content, and the latest updates.
Check us out on EarthSnap, a free app brought to you by Eric Ralls and Earth.com.
—–
