Breakfast cereal bowls, deli sandwiches, pizza dinners, soups, yogurt plates. Most people do not eat from a blank slate, they eat from habit. That is part of what makes nutrition advice so hard to follow. It is also part of what a new artificial intelligence system tried to solve.

Rather than designing ideal meals from scratch, researchers at the University of California, Davis built a model around the meals people already eat. The goal was simple: keep meals recognizable. Then, see whether a very small number of ingredient swaps could make them better aligned with dietary targets. The researchers also looked for ways to make the meals less expensive at the same time.

The answer, at least in a computational test, was yes.

Using national U.S. dietary survey data, Trevor Chan and Ilias Tagkopoulos developed a framework that generated realistic breakfast, lunch, and dinner meals based on common eating patterns. Then, the system searched for one-, two-, or three-item substitutions that improved nutrition. In the study, published in PLOS Digital Health, those limited changes improved nutritional quality by about 10%. They also cut modeled meal costs by 22% to 34%.

That combination matters because diet advice often breaks down at the point where science meets real life. People may know they should eat better, but broad guidance about protein, fiber, sodium, or saturated fat does not always translate into what to buy. In addition, people do not always know what to cook or swap on a Tuesday night.

The team drew on the USDA’s What We Eat in America survey, using six survey waves from 2013 to 2020. Their dataset included 135,491 meals logged by 55,228 adults. From there, the researchers grouped meals into recognizable breakfast, lunch, and dinner patterns, which they called archetypes.

Those archetypes captured familiar meal styles across the American diet, including cereal breakfasts, sandwich lunches, pizza dinners, soups, yogurt-based meals, and snack-like plates built around breads and spreads. In total, the researchers retained 34 meal clusters: 12 for breakfast, 11 for lunch, and 11 for dinner.

The point of clustering was not just organization. It gave the model a way to work within the logic of real eating habits. A lunch plate was compared with other lunches, not with an abstract idea of perfection. A cereal bowl stayed a cereal bowl.

The meal generator itself used a conditional variational autoencoder, a kind of machine learning model designed to produce plausible combinations rather than fixed outputs. It was trained on curated meals and then paired with a portion-adjustment system. That system nudged serving sizes toward USDA nutrient targets while trying to preserve the core makeup of each meal.

Compared with real meals from the same pattern, the AI-generated meals were 47% closer to USDA nutritional targets overall. That improvement was seen in 33 of the 34 meal clusters the team studied.

Lunch showed the biggest median reduction in deviation from targets, at 52.1%. Breakfast improved by 43.2%, and dinner by 46.0%. The researchers also found better adequacy across several nutrients, including fiber, protein, potassium, and vitamin C.

Not everything improved. Sodium deviation increased for generated lunches and dinners, a reminder that even when a model moves meals closer to guidelines overall, some nutrients can still lag or worsen.

The substitution step may be the study’s most practical piece.

After generating meals, the system looked for the fewest ingredient changes needed to improve them further. It compared meals with similar real-world alternatives, requiring roughly similar energy and meal size. Also, it tested simple one-item replacements within the same food category.

Across 19,013 generated meals, 8,337 had at least one feasible substitution candidate. Those are the meals included in the cost-benefit analysis.

At the study’s selected operating points, one-item swaps produced a 5.2% nutrition gain and 22.0% cost savings. Two-item swaps raised the nutrition gain to 8.1% and cost savings to 30.2%. Three-item swaps reached a 10.2% gain with 33.8% savings.

The most common patterns were not radical. The system often added vegetables or legumes and removed or replaced higher-sodium, processed items. In other words, it did not usually demand a total redesign of the plate.

That may help explain why the authors see promise in a minimal-change approach. “Dietary guidelines often tell people what a healthy diet should look like, but they do not always show how to get there from the meals people already eat,” Chan and Tagkopoulos said.

“Our study shows that it is possible to translate dietary standards into practical meal-level changes by identifying a small number of ingredient substitutions that can make meals healthier and cost-effective, while keeping them recognizable…[w]hat we found most interesting is that improving meals does not necessarily require a complete redesign. In many cases, targeted substitutions may be enough to move a meal closer to dietary recommendations, which could make healthy eating feel more practical and achievable,” he continued.

They put the point even more plainly in a second statement: “Healthier eating does not have to mean giving up the meals people already enjoy. With AI, we can identify small ingredient substitutions that preserve taste, while are better for our health and our pocket.”

The researchers also compared their specialized system with GPT-4o, using 3,400 meals across the 34 meal clusters. Their model performed better on most standards-based nutrition measures, especially macronutrient balance.

Only 11.9% of GPT-4o’s meals met AMDR macronutrient standards, compared with 18.9% for the UC Davis framework. GPT-4o also tended to generate meals that were higher in fat and lower in carbohydrates than the study’s model. GPT-4o did produce more diverse meals, but diversity was not the main goal here.

That comparison helps underline the paper’s broader argument: nutrition tools may work better when they are built with domain constraints instead of relying on a general chatbot to freestyle meal advice.

Still, the paper is careful about what the system has and has not shown.

The entire evaluation was computational. The researchers did not test whether real people would want these swaps, understand them, follow them, or keep using them. They did not measure clinical outcomes. They also note that the underlying diet data come from self-reported food intake, which can include underreporting, misreported portion sizes, and other biases.

The cost model, meanwhile, was based on a portion-based restaurant pricing framework at one point in time, not real grocery bills across different regions.

So the results show what the system can optimize on paper, not what it can yet change in kitchens, cafeterias, clinics, or shopping carts.

The study points toward a different kind of nutrition guidance, one that starts with the meals people already eat and suggests only a few targeted changes. That could make the advice feel more usable for consumer apps, public health programs, and, eventually, clinician-guided tools.

The appeal is not just better nutrient alignment, but lower friction: swap one or two items, spend less money, keep the meal familiar. The researchers stress that real-world testing still has to happen. This is especially important around usability, allergies, cultural fit, and whether people actually follow the advice.

But the work suggests that healthier eating may become more practical when the goal is not a perfect plate, but a slightly better version of the one already in front of you.