The Category Problem: What Counts as Ultra-Processed Food?
The central problem with ultra-processed food is not that the term is meaningless. It is that the term names a broad processing-based category, while the health question often treats it as if it were one substance. To understand what a reported UPF percentage or health association means, we need to see what the classification captures, which different foods it groups together, and which potentially important features it leaves entangled. That distinction is the difference between interpreting the evidence and merely reacting to the label.
The most widely used framework in the supplied research is NOVA. In broad terms, it places foods in a category of industrial formulations made from substances extracted or modified from foods, often containing few or no whole-food ingredients and additives used to improve shelf life, texture, appearance, or taste. The classification is concerned with the nature and purpose of processing, not only with the amount of sugar, salt, fat, or calories in a product.
That makes the category relevant to questions that a nutrient table may not answer. Industrial processing can change the physical form of a food, how quickly it can be eaten, how long it lasts, and how much preparation it requires. It can also change the commercial environment around it: packaged foods are easier to store, transport, portion, advertise, and consume away from a kitchen. But these features do not move together in a perfectly predictable package. A processing label is a useful starting point for inquiry, not a complete description of a diet.
A category built from unlike products
Consider a few of the product groups that appear repeatedly in the studies: sugar-sweetened beverages; sweet or salty packaged snacks; processed meats; industrial breads and breakfast cereals; ready-to-eat or ready-to-heat dishes; pizzas and instant noodles; and flavored or artificially sweetened dairy products. These foods can all be classified as ultra-processed in a study, yet they differ in obvious ways. A drink has no chewing requirement. A processed meat may provide protein but also contain substantial salt. A packaged bread may be used as one component of an otherwise varied meal. A ready meal may differ from a snack not only in ingredients but in portion, timing, and whether it replaces cooking.
The point is not that one of these examples is harmless or that another is uniquely responsible for weight gain. The point is that “UPF” does not tell us enough to infer a common effect at the product level. A category-level result can indicate that diets containing more of this collection of foods are associated with an outcome. It cannot, without additional evidence, tell us that every item in the collection contributes equally, through the same pathway, or independently of its nutrient profile and eating context.
This is also why food classifications and popular language can drift apart. In ordinary conversation, “processed food” may mean almost anything that has been cooked, canned, frozen, milled, or preserved. The NOVA-style UPF category is narrower and is intended to distinguish industrial formulations from less extensively processed foods. At the same time, the category is not identical to “junk food,” and “factory-made” is not a sufficient scientific definition. Processing can be used to make a product energy-dense and highly palatable, but the label alone does not reveal its full nutrient composition, portion, role in a meal, or likely substitute.
The USDA review shows how this breadth enters research. Among children and adolescents, studies classified UPF in different ways: some used NOVA directly, while others built dietary patterns through factor, cluster, or latent-class methods. Those latter approaches might identify a pattern dominated by foods typically consumed in ultra-processed versions, but they were not necessarily designed to measure processing as the main exposure. Studies also expressed exposure in different units, including grams, calories, percentages of energy, servings, or categories of adherence. A “high-UPF” group in one study therefore need not represent the same diet as a “high-UPF” group in another.
The adult evidence has the same problem at a larger scale. In studies reporting the information, UPF contributed anywhere from roughly 5% to 74% of total energy. That is not a single dose range in the ordinary sense; it is a collection of different populations, food supplies, measurement methods, and definitions. A person’s percentage can rise because they consume more sweet drinks, more packaged breads, more processed meat, or a different combination altogether. The number is meaningful only alongside the foods and comparator that produced it.
A classification is therefore an interface between the real diet and the analysis. The diet contains products, ingredients, textures, prices, routines, and substitutions. The study converts those details into an exposure variable. What crosses that interface depends on the rules used to classify foods and the tool used to record intake. If the interface is coarse or inconsistently applied, the resulting estimate may blur meaningful differences before the statistical analysis even begins.
The diagram’s important relationship is not the list of product classes. It is the direction of the arrows. The category groups products, and those products carry several features at once. A study can ask whether the grouped exposure relates to weight or another outcome, but the grouping does not automatically identify which feature is doing the work.
What the label does—and how to read it
One contested question is whether the health associations attributed to UPF are simply the familiar effects of an unfavorable nutrient profile. Many UPFs are energy-dense and rich in refined starches, free sugars, saturated fats, or salt, while being lower in fiber or micronutrients. If people eat more of them, they may consume more total energy or displace foods with different nutrient and fiber content. A nutrient-based explanation is therefore plausible and important.
But “plausible” does not mean “settled.” The Frontiers review notes that associations in prospective research often remained after adjustment for diet quality, which suggests that processing-related features might contribute beyond nutrient composition. That observation does not prove an independent processing effect. Dietary-quality adjustment can itself be imperfect, and processing, nutrient profile, texture, and eating behavior are often correlated. It tells us that the question should not be closed by assuming in advance that one label or the other must explain everything.
Texture is a good example of the overlap. A soft or finely structured food may require less chewing and can be consumed quickly. Faster eating may give satiety signals less time to catch up with intake. A sweet drink can deliver energy without the physical bulk of a solid meal. A ready-to-eat product can reduce the friction involved in obtaining food, which may change when and how much someone eats. None of these possibilities is encoded completely by the word “ultra-processed,” and none applies identically to every food in the category.
The same is true of additives and food structure. The category may include emulsifiers, flavorings, bulking agents, preservatives, or other ingredients, but the presence of an additive does not establish that it causes an observed outcome. Processing can also generate compounds or alter the matrix of a food, yet the health relevance of each change depends on the product, dose, exposure route, and duration. The supplied randomized review discusses these as possible pathways and emphasizes that no single mechanism has been established for all UPFs and all health outcomes.
Substitution makes the interpretation more difficult still. A dietary pattern with more UPF is not only an addition to the diet; it may replace fruit, vegetables, legumes, home-prepared food, or another product. The health association could reflect some combination of what is consumed, what is displaced, and what the surrounding lifestyle makes possible. The same packaged product may function differently when it is an occasional part of a varied diet than when it becomes a major source of daily energy. The classification records the product’s processing status, but not the full counterfactual meal that would have been eaten instead.
Price and convenience are part of this interface as well. Foods that last longer and require little preparation can be valuable where time, equipment, transport, or money is limited. A recommendation to reduce UPF may be easier to follow for someone with reliable access to affordable alternatives than for someone whose local food environment offers few practical substitutes. That does not change the biological question, but it changes the meaning of exposure, replacement, and intervention in real life. A study or policy that ignores those conditions may misread both behavior and feasibility.
Classification has two linked sources of uncertainty: the rule used to assign a product to a processing group and the quality of the dietary information fed into that rule. A questionnaire may record “breakfast cereal,” “yogurt,” or “bread” without capturing the formulation details that distinguish one product from another. Even when the ingredients are known, evaluators may not agree perfectly about how a mixed or reformulated product should be classified. That uncertainty can move people or foods between exposure groups, making genuinely different diets look more alike. It can weaken an association, distort its size, or make comparisons across studies look more consistent or more different than they really are. It does not, by itself, prove that a finding is false.
There is also a time problem. A person’s diet is not a fixed object, but many studies reduce it to one baseline assessment. Someone classified as having a high-UPF pattern may later change jobs, income, household responsibilities, health, or access to food, while the analysis continues to treat the original classification as representative. A data-driven pattern can add another layer: researchers may discover that certain foods travel together statistically, then give that pattern a name that sounds like a direct processing measure even though the original method was not designed to isolate UPF. The number produced by the study is therefore a translation of a changing diet, not a direct photograph of a single exposure.
The most important methodological boundary follows from this heterogeneity: the USDA review found no included study that directly compared the same types of foods in ultra-processed and non-ultra-processed versions. Such a comparison would not solve every problem, but it could make the processing question more specific. For example, it could ask whether two foods with similar energy density and broad nutrient composition produce different outcomes because their structures, textures, additives, or preparation histories differ. Without that kind of comparison, many studies evaluate dietary packages rather than processing as an isolated variable.
When a study reports that one group consumed more or less UPF, five questions prevent the number from doing more work than it can support.
First, what foods were counted? The study should describe the products or food groups, not only provide a label. If the high-exposure pattern is dominated by sweetened drinks, it is not interchangeable with one dominated by packaged breads or processed meats.
Second, which classification system and rules were used? NOVA is common, but the USDA review notes that studies also used data-driven patterns that were only indirectly related to UPF. A pattern named “Western,” “snacks,” or “energy-dense foods” may contain many UPFs without being a direct measure of their amount.
Third, what is the unit? Grams per day, calories, percentage of energy, servings, and a score across categories answer different questions. A percentage can rise either because UPF intake increased or because the rest of the diet decreased. A score can conceal which foods drive it.
Fourth, what was the comparator? “Lower UPF” might mean a different level of the same dietary pattern, a pattern with few or no foods in the category, usual intake, or a fully designed alternative diet. The contrast determines what the result can mean.
Fifth, what features travel with the category? Ask about energy density, sugar, salt, fat, fiber, texture, portion size, eating speed, access, and substitution. These are not reasons to dismiss a UPF association. They are the variables needed to understand it.
The practical conclusion is modest but powerful. Ultra-processed food is a real research category, and the repeated associations found in the evidence deserve attention. But the category is too heterogeneous to function as a complete biological explanation. Before asking whether higher UPF intake predicts greater adiposity, we need to know what “higher” meant in that study and which foods, nutrients, behaviors, and contexts were carried along with it. The next question is what pattern appears when researchers follow people over time—and how much confidence that pattern should receive.