Is "ultra-processed" a useful category? What NOVA gets right, and where it falls apart

A hundred and fifty-nine French nutrition specialists were handed a list of 120 supermarket products, complete with ingredient labels, and asked to do one thing: sort each item into the four NOVA groups. Which of these is ultra-processed? The people doing the sorting had the training, the definition and the ingredient list in front of them.

They agreed with each other at a Fleiss’ kappa of 0.32.

Bit nerdy, but stay with me, because that number is the whole post. Kappa asks how much of the available agreement — the gap between what you’d expect from luck and perfect consensus — the raters actually closed. At 1.0 they agree completely. At 0 they do no better than their own answer-frequencies would produce by chance, which with four categories and no even split between them is a subtler baseline than a coin toss, and worth not glossing: 0.32 does not mean raw agreement was near-random, it means they closed about a third of the distance to consensus. On the scale everyone uses, that is “fair” — the tier you would be quietly embarrassed to publish if you were validating a diagnostic test. Braesco and colleagues, who ran that survey and published it in the European Journal of Clinical Nutrition in 2022, got 0.34 on a second list of 111 generic foods without ingredient information. Roughly the same answer twice.

So we have a food category that half a decade of headlines, several national dietary guidelines and a great many books have been built on, and the experts can’t reliably say what’s in it. That seems worth sitting with.

I want to be careful here, because this is exactly the sort of finding that gets weaponised by someone who would very much like you to keep buying their product. This is not a post arguing that ultra-processed food is fine. It isn’t. It is a post about what the label can and can’t do — and my honest read is that “ultra-processed” is a genuinely good heuristic for filling a trolley and a genuinely bad variable for judging a food.

The claim, as everyone states it

The strong version goes like this. Nutrition science spent fifty years arguing about nutrients — fat, then cholesterol, then sugar — and got nowhere, because the real problem was never any single nutrient. It was industrial formulation itself. The NOVA system, developed by Carlos Monteiro’s group at the University of São Paulo, cuts through all of it by sorting food on how it was made rather than what’s in it. Group 4 — ultra-processed — is the one that matters, and the evidence that it predicts death, diabetes and heart disease is now overwhelming.

The practical instruction that follows is beautifully simple: don’t eat group 4. It requires no arithmetic, no app and no label-reading beyond “is this list long and weird”. Which is a large part of why it caught on.

If you want the full version of that case, with the cohorts and the mechanisms laid out properly, I wrote it up in the piece on ultra-processed foods and longevity. It is still what I would hand someone who has never encountered the idea. This post is the argument with it.

Where the category came from

NOVA was published in 2009 and refined over the following decade, and it was designed for population-level work — comparing national diets, tracking what happens to a country’s food supply over thirty years. At that job it is genuinely excellent. When you want to say “the UK food supply is now more than half industrial formulations”, NOVA gives you a way to say it.

The trouble started when a population-level instrument got promoted to a consumer-level rule. Michael Gibney set this out in Current Developments in Nutrition in 2019: the definition of group 4 has drifted repeatedly since inception, and because the written definition is hard to apply, NOVA’s own authors and others have published example lists instead — lists which have themselves varied considerably between publications. Gibney’s sharpest example is that commercially produced infant formula is classified as ultra-processed and, on the logic of the framework, to be avoided. Whatever you think of formula, “avoid it” is not a conclusion anybody arrived at by looking at outcomes in infants. It falls out of the classification.

That is the tell. When a category starts generating recommendations nobody would defend on the evidence, the category is doing work the evidence isn’t.

The reproducibility problem, in practice

The kappa figure isn’t an abstraction. Try it yourself on four items.

Tinned baked beans in tomato sauce. Wholemeal supermarket bread. A tub of plain skyr with added live cultures. A protein shake made of milk protein isolate, cocoa and sucralose.

Assuming the bread is the usual supermarket sort, it and the shake are group 4 — both contain emulsifiers or isolates you would not keep in a kitchen. The beans are usually group 3, the skyr group 1 or 3 depending on who’s asked. And note what that first assumption is doing: a bakery loaf of flour, water, yeast and salt is group 3, so the same wholemeal bread moves tier on the strength of an emulsifier, without changing much of anything a nutritionist would measure. Meanwhile the bread and the shake are, for most people, the least harmful things in that list, and I’d cheerfully argue that wholemeal bread of either tier has better evidence behind it than the beans. And a home-baked cake made of butter, sugar and white flour is group 3, and therefore in the same tier as the tinned beans.

None of that is a gotcha. It is the direct consequence of sorting food by manufacturing route rather than composition. The system is behaving exactly as designed; the design just doesn’t map onto “will this help me age well”.

What happens when the cohorts stop treating group 4 as one thing

Here is where the evidence gets genuinely interesting, because the best studies have already started doing the disaggregation themselves.

The umbrella review most people cite — Lane and colleagues in The BMJ, 2024, pooling 45 meta-analyses — reported roughly a 21% higher risk of all-cause mortality at the highest ultra-processed intakes. That’s the headline number, and it’s the one that made the news.

Then Fang and colleagues published a 30-year analysis of the Nurses’ Health Study and the Health Professionals Follow-up Study in The BMJ the same year: 74,563 women followed from 1984 and 39,501 men from 1986, all free of cancer, cardiovascular disease and diabetes at baseline. Comparing the highest quartile of ultra-processed intake with the lowest, all-cause mortality was about 4% higher.

Four percent. Over thirty years, in the best-characterised nutrition cohorts on earth, with repeated dietary assessment rather than a single food-frequency questionnaire. That is not nothing, but it is a fifth of the pooled figure, and it is the kind of effect size that a bit of residual confounding could plausibly account for on its own.

More useful than the headline was what happened when they broke group 4 apart. The association concentrated in ready-to-eat meat, poultry and seafood products, sugar-sweetened and artificially sweetened beverages, dairy-based desserts, and ultra-processed breakfast foods. And overall diet quality was inversely associated with mortality regardless of how much ultra-processed food someone ate — which is about as direct a statement as an epidemiologist will make that the processing variable is not the one carrying the load.

The European data says the same. Cordova and colleagues, in The Lancet Regional Health – Europe in 2023, followed 266,666 adults across seven countries in the EPIC cohort and looked at multimorbidity — developing two or more of cancer, cardiovascular disease and type 2 diabetes. Total ultra-processed intake was associated with higher risk, and among subgroups the associations were most notable for animal-based products and for sweetened beverages.

Two enormous, independent, differently-designed studies, and both land on the same short list: processed meat, and drinks with sweeteners in them. You did not need NOVA to arrive at that list. Dietary guidelines were saying it in the 1990s.

The two trials that bracket what processing actually buys

The observational data can only ever tell you about association, so the causal question rests on two feeding trials — and read together they’re much more informative than either alone.

Hall and colleagues, Cell Metabolism, 2019. Twenty adults lived in a metabolic ward at the US National Institutes of Health and ate an ultra-processed diet for two weeks and an unprocessed one for two weeks, in random order. The meals were matched on presented calories, energy density, macronutrients, sugar, sodium and fibre. People ate 508 ± 106 kcal a day more on the ultra-processed arm and gained weight; on the unprocessed arm they lost it. This is the single most-cited result in the field and it deserves to be, because it isolated processing about as cleanly as anyone has managed.

Dicken and colleagues, Nature Medicine, 2025. Fifty-five adults with a BMI of 25–40 who already got at least half their calories from ultra-processed food were given two eight-week ad-libitum diets in a crossover, with a four-week washout. The crucial difference from Hall: both diets followed the UK Eatwell Guide. One was built from minimally processed foods, one from ultra-processed ones, and both hit national nutrition recommendations.

Both arms produced weight loss. The minimally processed diet produced −2.06% of bodyweight; the ultra-processed diet −1.05%. Fat-mass loss and food cravings also favoured the minimally processed arm. The cardiometabolic markers were a genuine mess in both directions — triglycerides fell further on minimally processed, LDL cholesterol fell further on ultra-processed — and there was a substantial order effect, with the difference much larger in the first period than the second. Several statisticians pointed that out at the time, correctly.

Put the two trials next to each other and the honest thing to say is that they measure different quantities, so you cannot subtract one from the other. Hall measured energy intake — about 500 kcal a day — over two weeks, in weight-stable adults living in a metabolic ward with unlimited access to whichever diet they were on. Dicken measured bodyweight over eight weeks, in free-living adults with a BMI of 25–40 who were already eating half their calories as ultra-processed food. Different outcome, different population, different duration, different setting. Anyone quoting “500 kcal” and “one percentage point” as two points on the same scale — and I nearly did — is comparing kilocalories per day with a percentage of bodyweight, which is not a comparison at all.

What survives is narrower and still worth having. The trial that let ultra-processed food behave the way it does on a supermarket shelf found a large effect on how much people ate. The trial that made both arms hit national nutrition guidance found a small effect on what they weighed. Composition is not the only thing that differs between those two designs — but it is the thing the second trial deliberately held constant, and holding it constant is where the effect got small. That is a claim about direction and about context. It is not a measured size for “processing itself”, and nobody has one.

The strongest counter-argument

The honest steelman, and it’s a good one: a heuristic doesn’t have to be precise to be useful. “Ultra-processed” moved more people toward real food in five years than “reduce your saturated fat intake” managed in forty, and it did so without requiring anybody to read a nutrition panel. Public health advice is judged on what people actually do with it, not on inter-rater reliability. By that standard NOVA has been a roaring success, and Dicken’s trial shows there is a real, measurable benefit sitting underneath the slogan even when you control the nutrition side.

There’s a second, subtler defence. NOVA might be capturing something that composition tables genuinely miss — additive effects on the gut lining, the eating-rate consequences of soft textures, the cumulative effect of hyper-palatable formulation on appetite regulation. Composition-based systems can’t see any of that, and Hall’s 500 kcal has to be coming from somewhere.

I find both of those persuasive. They just don’t rescue the label as a way to judge one food on one shelf. They argue for it as a shopping posture.

The honest verdict: a good heuristic, a bad variable

“Ultra-processed” works as an instruction and fails as a measurement.

As an instruction — fewer industrial formulations, more things that look like food — it is one of the better pieces of nutrition advice in circulation, and I follow it. As a measurement, it has fair inter-rater reliability at best, a definition that has moved repeatedly, a group 4 that contains both a supermarket sponge cake and a mass-produced wholemeal loaf, and a mortality signal that shrinks to 4% and relocates into three named subgroups the moment anyone looks properly.

The failure mode worth avoiding is treating the label as a verdict on a specific item. That’s how you end up buying an expensive “clean” biscuit that is nutritionally identical to a cheap one, or refusing tinned lentils, or feeling virtuous about a group 3 cake. The category was never designed to answer that question and it doesn’t.

What to use instead: three questions that beat the label

When I’m actually standing in the Viktualienmarkt or, more honestly, in an Edeka at half seven on a Tuesday, these are what I use.

1. How fast can I eat it? Forde, Mars and de Graaf, writing in Current Developments in Nutrition in 2020, pooled 327 foods and measured energy intake rate — kcal consumed per minute of eating. It rose across the NOVA groups, from an average 35.5 kcal/min for unprocessed foods to 53.7 for processed and 69.4 for ultra-processed. But the overlap was substantial: plenty of unprocessed foods exceeded 100 kcal/min. Their argument is that energy density and eating rate, not processing per se, are what drive overconsumption — and that these are measurable properties of the actual food in front of you rather than a category judgement about its factory. Soft, dense, fast: eat with care. Bulky, chewy, slow: less so.

2. Is it on the short list? Processed meat, sweetened drinks (sugar or artificial — both showed up in Fang and in EPIC), sugary breakfast foods, dairy-based desserts. That’s where the signal actually lives across both big cohorts. If your ultra-processed intake is mostly supermarket wholemeal bread and fortified yoghurt, you are having a different conversation from someone whose intake is mostly bacon and Cola Zero.

3. What is it displacing? This is the one nobody asks. A protein-fortified yoghurt that replaces a pastry is a good trade even though both are group 4. Fibre is the clearest case — most people are nowhere near adequate intake, and dietary fibre has some of the most consistent mortality data in nutrition, so a high-fibre industrial cereal beats a low-fibre artisanal one on every axis that has been measured. The same logic runs through the wider longevity-diet evidence: what a dietary pattern contains has consistently outperformed what it excludes.

Those three get you most of what NOVA gets you, plus the resolution to tell a mass-produced wholemeal loaf from a Swiss roll — which NOVA, having filed both in group 4, cannot do on its own terms.

The short version

The mortality association is real, replicated and pointing the right way. The category carrying it is imprecise enough that trained specialists holding the ingredient list agree only fairly on what belongs in it, heterogeneous enough to hold both a sponge cake and a mass-produced wholemeal loaf, and — when the two best-resourced cohorts disaggregate it — mostly a proxy for processed meat and sweetened drinks. The two feeding trials both favour eating less of the stuff, and they measure different things in different populations, so they establish the direction of the causal claim without pinning down its size.

So: keep using it to fill a trolley. Stop using it to convict a food. And if you want a single number to actually watch while you’re changing what you eat, post-meal blood glucose is a far more direct read on how a specific food behaves in your specific body than any classification system will ever be.

Which is, I’ll admit, a slightly deflating conclusion for a framework that promised to finally settle nutrition. Turns out it’s still mostly “eat food, mostly plants, not too fast”. Fair enough.