Ultra-Processed Food: What the Research Actually Shows

The NOVA classification system, developed by Carlos Monteiro and colleagues at the University of São Paulo, categorizes foods by their degree of industrial processing rather than nutrient content:

The defining characteristic of Group 4 (UPFs) is the use of food substances extracted or derived from foods, and additives that serve industrial functions: emulsifiers, stabilizers, colors, flavors, texture agents. These are not found in home cooking and are not used to “preserve” food but to create desirable sensory properties (texture, flavor, color) and extend shelf life.

According to Steele et al. (2016, American Journal of Preventive Medicine), using NHANES 2009–2010 data: UPFs provided 57.9% of total energy intake and 89.7% of added sugar intake in the U.S. adult diet. More recent estimates suggest this has risen to approximately 60%. By contrast, UPF consumption is lower in most European nations (30–50% of energy intake) and substantially lower in many middle-income countries.

The strongest experimental evidence for UPF effects on caloric intake comes from a randomized controlled crossover trial by Kevin Hall and colleagues at the National Institutes of Health (Cell Metabolism, 2019). This is the highest-quality design available for diet research because it controls for confounders that plague observational studies.

The critical finding is that despite being offered matched nutritional content, participants ate significantly more calories on UPF diets. This suggests that processing characteristics beyond macronutrients—texture, palatability, eating rate, texture disruption—drive overconsumption independently of caloric density. The study was small (n=20) and short (2 weeks each), limiting generalizability, but it is the strongest causal evidence available for UPF effects on food intake.

A follow-up RCT (Dicken et al., BMJ 2025) comparing weight loss on minimally processed vs. UPF-matched diets over 12 weeks (n=36) found that participants on the minimally processed diet lost significantly more weight (−2.06% body weight) than those on the matched UPF diet (−1.05%), despite matched macronutrients and calorie targets. This extends the Hall findings to a longer-term, weight-loss context.

Fiolet et al. (BMJ, 2018) found a 12% increase in cardiovascular disease risk per 10% increase in UPF dietary share (n=105,159, NutriNet-Santé cohort, 5-year follow-up). Srour et al. (2020, Lancet) found HR 1.11 for cardiovascular disease and HR 1.17 for coronary heart disease, highest vs. lowest UPF consumers. Effect sizes are modest but consistent across multiple large cohorts.

All cohort studies face the same fundamental challenge: UPF consumption correlates with other markers of unhealthy lifestyle (lower income, lower education, less exercise, worse sleep, higher stress), making it difficult to isolate UPF effects even with extensive statistical adjustment. The “healthy user bias” problem is the central limitation. Multiple well-controlled cohorts adjusting for BMI, income, education, physical activity, smoking, and overall dietary quality continue to find associations—but residual confounding cannot be fully excluded from observational data alone. This is why the Hall 2019 RCT, despite its small sample, is particularly valuable: it controls these confounders experimentally.

Several non-mutually-exclusive mechanisms are proposed, with varying levels of evidence:

The NOVA UPF category is broad. It includes both Coca-Cola and whole-grain bread made with dough conditioners. Cohort studies that break out UPF subcategories consistently find heterogeneity:

This heterogeneity has important implications: treating all UPFs as equivalent may be an oversimplification. The harm signal in aggregate UPF analyses is likely driven disproportionately by processed meats, SSBs, and packaged snacks—not by whole-grain breads or low-fat dairy products that happen to qualify as “ultra-processed” under NOVA.

The Hall 2019 RCT is the key piece of evidence: an inpatient, controlled experiment where matched nutritional content did not prevent +508 kcal/day overconsumption on UPF diets. This is causal evidence that processing characteristics—independent of macronutrient content—drive overconsumption. Combined with consistent associations across 10+ large cohort studies (combined n >1M) showing elevated mortality and disease risk, and the plausible mechanisms (hyperpalatability, matrix disruption, displacement), the evidence for harm is substantial.

The displacement effect is also important even without direct UPF toxicity: when 58% of calories come from UPFs, less room exists for fruits, vegetables, legumes, and whole grains with well-established protective effects. The harm may be partially indirect but is nonetheless real.

The observational evidence cannot establish causation due to confounding. People who eat more UPFs also tend to be lower income, less educated, less physically active, under more stress, and sleep less—all independent risk factors for the outcomes being studied. Even well-adjusted models cannot fully exclude these correlates.

The NOVA classification lumps foods with very different nutritional profiles and health effects into the same category. Yogurt and hot dogs are not comparable health risks. Policy based on broad “ultra-processed” categorization may be less effective than targeted advice on the highest-risk subcategories (processed meat, SSBs).

The RCT evidence (Hall 2019) is compelling but small (n=20) and short (2 weeks). Scaling these effects to long-term population outcomes requires assumptions the data cannot directly support. A larger, longer-term RCT is needed but extremely difficult to conduct due to the impracticality of years-long controlled feeding studies.