The problem with "Ultra-processed food"

The concept of "ultra-processed food" has gone from one Brazilian commentary in 2009 to a global policy framework in fifteen years. The term shows up in national dietary guidelines in Brazil, Uruguay, Ecuador, Chile, Peru, Mexico, Belgium, Israel, Malaysia, and Maldives (Bartholmae et al. 2023), in France's PNNS 2018-2022 plan with a 20% reduction target, and in the Lancet's November 2025 three-paper series calling for global policy reform. While sometimes it can be good that policy is quick to catch up to science, going too fast comes with a risk when the underlying evidence is weak.

I think the case that ultra-processing itself drives disease is weaker than the headlines suggest, and I want to lay out why. There are four problems worth taking seriously.

Problem 1: Definition

The NOVA classification splits food into four groups based on "the extent and purpose of processing" (Monteiro et al., 2019): Group 1 is fresh or minimally processed food, Group 2 is culinary ingredients like oil and salt, Group 3 is processed food made by combining Groups 1 and 2 (canned vegetables, simple cheeses, freshly baked bread), and Group 4, ultra-processed, is defined as "formulations of ingredients, mostly of exclusive industrial use" that result from "a series of industrial processes." Protein bars, oat milk, whole-wheat supermarket bread, plant-based burgers, infant formula, and candy are all ultra-processed foods in this description.

However, this description leaves a lot of room for interpretation. Braesco et al. (2022) recruited 159 evaluators to classify 120 marketed food products with full ingredient lists, and 177 evaluators to classify 111 generic food items without ingredient information. All evaluators were food scientists, nutritionists, dietitians, or food industry R&D specialists. Inter-rater agreement, measured with Fleiss' κ, came out at 0.32 and 0.34 respectively. κ is a statistic scaled so that 0 corresponds to chance-level agreement and 1 to perfect agreement, with 0.21–0.40 being in the "fair" bucket and 0.41–0.60 in "moderate." A κ of 0.32 isn't random, but for a classification system meant to determine dietary guidelines and taxation policy, I'd say it's way too low.

Only 3 of the 120 marketed foods got the same NOVA assignment from every evaluator, and only 1 of the 111 generic foods. About a third of marketed foods (40 of 120) and three-quarters of generic foods (84 of 111) were placed in all four NOVA groups by different raters; giving people the ingredient list barely improved agreement. These same evaluators reported "high" or "very high" confidence in their assignments while disagreeing with each other, which suggests the disagreements stem from ambiguity in the system. When the researchers checked the nutritional quality of foods that the majority of evaluators classified as NOVA-4, a substantial fraction landed in the healthiest Nutri-Score categories: 26% and 35% respectively for marketed foods, 18% and 32% for generic foods. NOVA, in other words, was not tracking nutrient quality reliably. Figure 1 visualizes the assignment distributions across both lists.

Gibney (2019) tracked the example lists and definitions used across Monteiro group publications and showed that which specific foods count as "ultra-processed" has shifted repeatedly over the years. If trained specialists and the original inventors of the NOVA system can't classify consistently, it is hard to maintain that NOVA is measuring a well-defined property of food.

Problem 2: Heterogeneity

Smoking is an example of a categorical exposure that produces a categorical effect. Cigarettes, cigars, pipes, and hand-rolled tobacco all raise lung cancer risk, and within "cigarettes," every brand raises risk in the same direction, because the underlying mechanism is consistent across the category. No matter how you slice the data, the sign of the effect stays the same. In this case, there's a good link between the category and the causal mechanism.

The UPF data, in comparison, looks nothing like this. Chen et al. (2023), pooling 32 years of follow-up across the three large US Nurses' Health and Health Professionals cohorts, found a hazard ratio of 1.46 for total UPF intake and type 2 diabetes. When they stratified by NOVA-4 subgroup, processed meats, refined breads, sauces and condiments, and sugar-sweetened beverages drove most of the risk, while ultra-processed whole-grain breads, breakfast cereals, packaged sweet and savory snacks, fruit-based products, and yogurt and dairy desserts all showed inverse associations. A 2025 dose-response meta-analysis pooling 12 cohorts (Kim et al., 2025, Diabetes Metab J) reproduced the same pattern: processed meats and sugar-sweetened beverages increased risk, but ultra-processed cereals, breads, and packaged savory snacks reduced it. Figure 2 plots all subgroup estimates on the same axis.

If ultra-processing itself were the causal mechanism as presented, you would expect every subgroup of Group 4 to show harm, with maybe some heterogeneity in magnitude rather than reversal in direction. Alternatively, the category is a convolution of several things at once, and the aggregate effect that gets reported in headlines is whatever weighted average falls out of whichever subgroups happen to dominate the typical diet in a given cohort, like high sodium foods being relatively common in UPFs.

Problem 3: Experiments

Almost all of the UPF literature is observational, and every major cohort uses food frequency questionnaires. Every cohort adjusts for smoking, income, education, and physical activity, but those adjustments can only do so much when the exposure measurement is noisy, the confounder set is incomplete, and statistical correction can't completely control for confounders. Lane et al.'s umbrella review that lead to news headline about "32 adverse outcomes" related to UPFs itself rates most of the evidence as low or very low certainty under GRADE.

The randomized trial evidence is scarce, in comparison. Aramburu et al. (2024) systematically searched for randomized UPF intervention trials up to April 2024 and found four in total: three educational interventions, plus Hall's 2019 controlled feeding trial. Pooled across 455 participants, 30 of 42 health-related outcomes showed no significant effect, and the only trial that did show consistent effects across energy intake, body weight, and cholesterol was Hall's. Hall et al. (2019) admitted 20 adults to the NIH metabolic ward for four weeks, randomized them to a UPF or unprocessed diet for two weeks each in crossover, with both diets matched for calories presented, macronutrients, fiber, sugar, sodium, and energy density. Crucially, participants ate as much as they wanted, and they took in about 500 kcal/day more on the UPF arm and gained 0.9 kg. While this is clearly a real finding, it's a finding about eating rate and ad libitum intake rather than about the health effects of equicaloric diets. The conclusion that highly palatable energy-dense food drives overeating is not new.

The second major UPF RCT (which came out after the previous review), Dicken et al. (2025) in Nature Medicine, is the most cited recent piece of evidence in favor of the category. It was a 2×2 crossover in 55 UK adults with overweight or obesity (BMI 25–40, habitual UPF intake ≥50% of daily kcal), randomized to two 8-week ad libitum diets that both followed the UK Eatwell Guide: one built from minimally processed foods (MPFs), and one built from UPFs. The primary endpoint was percent weight change. Both diets produced weight loss: −2.06% (95% CI −2.99, −1.13) on MPF and −1.05% (−1.98, −0.13) on UPF. The within-participant difference was 1.01 percentage points in favor of MPF, p = 0.024, corresponding to roughly a 1 kg additional loss over 8 weeks. Estimated calorie deficits were about 290 kcal/day on MPF versus 120 kcal/day on UPF.

There are several things worth talking about regarding this trial. First, the headline number depends on averaging across both crossover periods: the first-period difference was about 1.86 percentage points, but the second-period difference was close to zero, which a letter to Nature Medicine from Ludwig, Willett, and Putt (2025) argues makes the extrapolation to "9–13% versus 4–5% annual weight loss" unsupportable. Second, secondary cardiometabolic outcomes were broadly equal: the authors themselves write that the greater weight loss on MPF "did not translate into significant improvements in cardiometabolic risk factors over the UPF diet, except triglycerides," and that the UPF diet produced reductions in heart rate, fasting glucose, total cholesterol, and LDL-C that were similar in magnitude to MPF. The cardiometabolic result is that both Eatwell-compliant diets, ultra-processed or not, were healthier than what participants were eating before. Third, the most striking effect was on appetite control rather than physiology: participants on MPF reported a roughly two-fold greater improvement in overall craving control on the Control of Eating Questionnaire and a four-fold greater improvement in savory craving control. Fourth, 10 of 50 participants gained weight on each arm, which the authors attribute to compliance failures particularly in the second crossover period. Lead author Dicken's own framing: "not all ultra-processed foods are inherently unhealthy based on their nutritional profile."

The closest things to mechanistic-level RCTs in this literature are studies of specific additives. Chassaing et al. (2022, Gastroenterology) showed that carboxymethylcellulose, an emulsifier commonly found in UPFs, altered gut microbiota composition and the fecal metabolome in healthy adults over 11 days, with a subset of subjects showing microbiota encroachment into the inner mucus layer. These are mechanistic rather than outcome trials, and they support the idea that specific things found in some UPFs have biological effects, but they do not test the UPF category as a whole.

Adding all of this up, the experimental base for "UPFs cause 32 adverse health outcomes" amounts to: one n=20 weight trial, one n=55 weight trial showing a modest effect within an otherwise-healthy diet plus null cardiometabolic differences, three educational interventions that were mostly null, and a handful of additive-specific studies that test particular compounds rather than the category.

Problem 4: Mechanism

If something about how industrial food gets formulated does drive bad outcomes, the candidates worth taking seriously are properties of particular foods rather than of the NOVA category. For weight and overeating specifically, Kevin Hall's unpublished 2024 NIH follow-up trial (slides from his Imperial College London presentation, summarized by Marion Nestle) tries to factor UPF diets by two correlated properties: energy density and hyperpalatability (concentrated combinations of fat, sugar, and salt that co-occur rarely in whole foods). A UPF diet high in both properties drove about 1000 extra kcal/day versus a minimally processed control. A UPF diet reformulated to reduce both properties cut that excess to roughly 370 kcal/day, a 63% reduction in the overeating effect. Hall's interim conclusion was that "weight gain is not a necessary component of a highly ultra-processed diet".

For health outcomes beyond weight, the plausible causal nodes are similarly specific. The candidate list includes trans fats, excessive added sodium, excessive added sugar, certain emulsifiers, and nitrite preservatives in cured meats. Each of these is a property of particular foods, and none of them link directly to "ultra-processed" as a category.

One potential exception is the Maillard reaction – the reaction which creates the brown pigment after heating on some foods, through reactions between amino groups and carbonyls on sugars. If at high enough temperature, it can cause potentially harmful compounds, like acrylamide, to form, but this would almost exclusively drive potentially carcinogenic effects. On top of that, minimally processed foods are often is also heated similarly.

When considering specific causal reason for health, there's a lot of heterogeneity in the UPF group. Oat milk is NOVA-4 and has essentially none of them, while a bottle of soy sauce or a generous piece of cured ham can deliver several at once without being NOVA-4. This is why the category-level policy proposals are potentially harmful: if you tax ultra-processed food and the real drivers are energy density, hyperpalatability, sodium, added sugar, and a short list of additives, you capture some harmful products and miss others, ending up taxing soy milk while leaving butter untouched.

A bad category is still a bad category

UPF is doing the same job as a shortcut like "foods in shiny packaging are unhealthy." On average that might correlate with something true, because the shiny-package category would correlate with chips and candy. But the packaging is not the mechanism, and the category sweeps in things that don't belong (plenty of frozen vegetables and whole-grain bread come in shiny bags) while missing things that do (most red meat, deep-fried food, sugary drinks in bottles). A scientific paper would not propose shiny-package taxes, because the category doesn't map to the causal structure. UPF has more face validity, but the same problem is underneath due to a lack of plausible biological mechanism.

Why does it even matter, if it somewhat correlates with true problematic foods? The problem is that the UPF framework can be used to argue against healthy food. It is now used in mainstream coverage to argue against plant-based diets, which depend on fortified plant milks, (often) plant-based meats, and whole-grain breads that all sit in NOVA-4, even though the evidence on these foods is quite positive.

If we want to work on the global food environment, we need to target the actual biological drivers of disease and overeating; currently our best guesses are energy density, hyperpalatability, sugars, saturated fats, excess sodium, and other specific harmful substances. The danger of the ultra-processed panic isn't that it attacks junk food. It’s that it pretends the factory is the enemy, when the real enemy is the recipe.

The Winner’s Curse: Why nutritional findings keep shrinking

Throughout the 1980s, there was a set of studies that seemingly found something very interesting: people who ate more beta-carotene-rich fruits and vegetables appeared to get less lung cancer. The logic was appealing - beta-carotene is an antioxidant and a precursor to vitamin A, which would be a plausible way

Measured ThinkingBrent Visser