Healthy eating advice often starts with an ideal: more vegetables, more whole grains, less sodium, less added sugar. But for many people, the harder question is what that looks like in the middle of a regular week, with familiar meals, a grocery budget and no interest in starting over.
A new study suggests that smaller changes may be part of the answer. Researchers developed an artificial intelligence framework that used more than 135,000 U.S. meal records to identify simple food swaps that could make meals more nutritious and less expensive while keeping them close to what people already eat.
The study, published in PLOS Digital Health, is not a real-world test of whether people would like, buy or stick with the suggested swaps. Instead, it is a computational study showing how AI could help translate dietary guidelines into more practical meal-level changes.
That distinction matters. The findings do not prove that an app can automatically improve someone’s diet, and they do not show that people will choose the recommended foods in real life. But they point to a useful idea: healthy eating may feel more realistic when it begins with small substitutions, not complete meal makeovers.
The researchers used data from What We Eat in America, a national dietary survey, to analyze 135,491 meals logged by 55,228 adults. They identified common meal patterns for breakfast, lunch and dinner, then trained a generative AI model to create realistic meals that followed those patterns while adjusting serving sizes.
Next, the researchers tested whether the system could identify one, two or three ingredient substitutions that would improve nutrition and reduce modeled meal costs.
Compared with real meals in the same dietary pattern, the AI-generated meals were 47% closer to USDA nutritional targets, while still staying close to the overall meal type and flavors people actually reported eating. When the system applied one to three food swaps, nutritional quality improved by about 10%, while modeled meal costs fell by 22% to 34%.
The most common substitutions involved adding vegetables or legumes and replacing high-sodium or processed items.
That is what makes the study especially interesting from a behavior-change perspective. Many nutrition tools ask people to change too much at once. This framework was designed to preserve the basic structure of meals people already recognize, then make targeted substitutions that move the meal closer to dietary recommendations.
“Healthier eating does not have to mean giving up the meals people already enjoy,” the authors said.
That idea may resonate because the gap between knowing nutrition advice and using it is often wide. People may understand that beans, vegetables or lower-sodium foods are good choices, but still struggle with how to fit them into the meals they already cook, order or pack.
A swap-based approach could make that process feel less overwhelming. For example, a meal does not necessarily need to be rebuilt from scratch to become more nutritious. Adding a vegetable, replacing a processed side with beans or choosing a lower-sodium version of a familiar ingredient may be more manageable than adopting an entirely new eating pattern overnight.
The study also compared the trained model with GPT-4o, an unspecialized AI model. The specialized model produced meals that were closer to USDA guidelines for macronutrients, suggesting that nutrition-specific training and constraints may matter when AI tools are used for food recommendations.
That point is important because AI-generated nutrition advice can sound confident even when it is incomplete or unrealistic. A useful food recommendation tool needs more than creativity. It needs to account for nutrition standards, cost, serving sizes, meal patterns and whether a suggested meal still makes sense to the person eating it.
The authors said the framework could eventually support public health programs, consumer apps or clinical tools. But the next step is testing whether these kinds of recommendations work outside a computer model.
Real people make food choices based on taste, culture, time, cooking skills, food access, family preferences and what is available at the store. A swap may look nutritionally better and less expensive in a model but still fail if it does not fit someone’s life.
For now, the study offers a promising direction rather than a finished solution. AI may eventually help make healthy eating advice more specific, budget-aware and realistic. But the broader takeaway does not require an app: small, targeted changes can be a practical place to start.
As the authors put it, improving meals “does not necessarily require a complete redesign.”
This work was supported by the USDA-NIFA AI Institute for Next Generation Food Systems and by the NSF HDR: TRIPODS program. Trevor Chan received salary support from both grants.