“Processed” is not one exposure but a dozen. A food’s matrix can be broken open, additives mixed in, a grain stripped and refined, calories packed denser, a recipe tuned for palatability, a contaminant formed in the cooking — and folding all of that under one alarming word is the mistake this decision has to undo. Split the word into its aspects and its food categories first, and most of the fear reattaches to nameable things a shopper can already price. The category average that drives the headlines — the ultra-processed (UPF) grouping in Lane’s 2024 umbrella review — is a broad, consistent set of associations with mortality, cardiometabolic disease and common mental disorders. But it is observational throughout, low-to-very-low certainty, modest in size (relative risks around 1.2 to 1.66), confounded by overall diet quality and income, and the bias screens fire on the very outcomes it headlines. (Lane et al., 2024)
The one randomized foothold points the same way but names the mechanism. Hall’s 2019 inpatient trial matched an ultra-processed and an unprocessed diet on calories, energy density, macronutrients, sugar, sodium and fibre, fed them freely, and found the ultra-processed arm drove ~500 extra kcal/day and 0.9 kg of weight gain. Processing does independent work on how much people eat — but on a two-week surrogate, in 20 people, and it runs through properties you can name: higher energy density, a faster eating rate, protein dilution, not “processing magic.” (Hall et al., 2019) So the honest lever is the components you can price — energy density, how fast a food is eaten, protein, added sugar, sugary drinks, processed meat — not the NOVA label a shopper cannot reliably apply.
The rest is smaller or thinner. Hyper-palatability is now a measurable food property (Fazzino’s composition definition), distinct from ultra-processing and from energy density — but it has no outcome evidence yet, a candidate mechanism rather than a sized lever. The intake evidence points to eating rate, a mechanical channel with no hunger penalty, rather than reward or palatability; the reward story, including “food addiction,” is a contested construct with no patient-important outcome behind it. Additives and contaminants are where mechanism outruns human data — insufficient evidence, judged compound by compound, not “additive therefore harmful.” Processed meat and colorectal cancer is the one category carrying an evidenced, if modest, harm. And against the big rocks — smoking, obesity, inactivity — none of this is one; if those are unaddressed, no processing precision changes the next move.
Split “processed” before you judge it
Before you can judge a processed food, you have to say which sense of “processed” you mean. The word bundles at least six distinct aspects — a disrupted food matrix, added additives, refining or substrate change, higher energy density, engineered hyper-palatability, and processing contaminants — and, separately, a dozen food categories that a classification like NOVA sweeps together. A claim that fits one aspect need not fit another, and a claim about a category is an average over everything inside it.
An average over a heterogeneous mix describes no actual food. The diagnostic is sharp: when the variation within a category exceeds the variation between categories, the boundary carries no information, and the category-level estimate is not merely imprecise — it describes nothing on any plate.
The ultra-processed grouping is the paradigm case, because it is defined by manufacturing rather than composition. Within a single diabetes meta-analysis Lane includes, some ultra-processed subcategories — wholegrain breads, yoghurt, some packaged snacks, fruit-based products — were inversely associated with risk while others raised it, so the category pools foods that help and foods that harm. (Lane et al., 2024) NNR names the same defect from the other side: the same bread or yoghurt sits inside or outside the category depending on how it was manufactured, so even a perfectly informative boundary would be one a shopper cannot apply at the shelf. (Nordic Council of Ministers, 2023)
So the method for the whole question is to attach every claim to a specific aspect or a specific category, never to “processing” bare. Sometimes the boundary is load-bearing — sugar-sweetened beverages carry a sugar-and-weight signal in child cohorts that other sugars do not (Morenga et al., 2013). Sometimes it is decorative — refined versus whole grain is null on hard outcomes (Scientific Advisory Committee on Nutrition, 2015). You cannot tell which case you are in without splitting. The first question the split raises is the sharpest one: does the ultra-processed label carry any causal work of its own, beyond the sugar, salt, fat and energy density it travels with?
The category average is real, but confounded and low-certainty
The category average is real: broad, consistent, and drawn from a lot of people. Lane’s umbrella review pools 14 meta-analyses and 45 distinct analyses covering nearly 9.9 million people, and reports direct associations between ultra-processed food and 32 of 45 distinct pooled analyses (71%). (Lane et al., 2024) But it has to be read endpoint by endpoint, because the signal is not uniform. Cardiometabolic disease, mortality and common mental disorders are where it is strongest — CVD-related mortality RR 1.50 (1.37-1.63), all-cause mortality RR 1.21 (1.15-1.27), heart-disease mortality HR 1.66 (1.51-1.84), common mental disorder OR 1.53 (1.43-1.63), and type 2 diabetes RR 1.12 (1.11-1.13) per 10% of intake. Cancer is where it is weak to absent — breast, prostate and pancreatic cancer, CNS tumours and cancer-related mortality all fall in the “no evidence” class. (Lane et al., 2024)
The magnitudes are modest, and that bounds the decision. The strongest relative effects run about 1.2 to 1.66 — small-to-moderate against the baseline risks of these outcomes, not the order of a big rock. (Lane et al., 2024) And a strong association is not strong evidence of cause. Lane grades every analysis twice, and the two axes disagree: the single strongest association — CVD mortality, rated class I convincing on the credibility axis — is GRADE very low. Only 4 of 45 analyses reach GRADE moderate (the type-2-diabetes dose-response among them); 22 are low and 19 very low. (Lane et al., 2024) The honest one-line read is broad, consistent associations at low-to-very-low causal certainty.
Certainty is floored low because the evidence is observational throughout — by necessity. Lane found no pooled randomized trials to draw on; long-term trials feeding ultra-processed diets to hard endpoints like cancer or cardiovascular disease are ethically impossible, so GRADE starts this whole body of evidence at low quality, liftable only by a dose-response gradient or a large effect. (Lane et al., 2024) Underneath even that sits dietary measurement error, large enough to be the dominant consideration in reading almost any observational nutrition finding.
The bias screens fire on the headline outcomes, not the weak ones. Lane’s re-analysis found excess-significance bias in 9 of 28 analyses with enough studies to test (32%) — including all-cause mortality, obesity, metabolic syndrome, NAFLD and type 2 diabetes — and small-study effects in 5 of 28 (18%), including all-cause mortality, breast cancer and obesity. (Lane et al., 2024) A nominally convincing association can still carry these flags, and here the strongest ones do.
And the whole association is confounded in a way adjustment cannot fully remove. People who eat more ultra-processed food have poorer overall diets and lower socioeconomic position, both of which independently predict the same outcomes — NNR treats ultra-processed intake as a marker of diet quality and deprivation, not a clean exposure. (Nordic Council of Ministers, 2023) Lane’s defence — that adjusting for diet-quality patterns leaves the associations standing — does not close the gap, because adjusting for a pattern is not the same as matching the sugar, salt, fat and fibre profile that makes a food ultra-processed and nutrient-poor at once. (Lane et al., 2024) An association this confounded needs a design that holds composition fixed — and the newest outcome arm, cognition, shows just how much of the association the confounding can swallow.
Dementia and cognition: the signal thins on adjustment
Add dementia to the outcome list and it behaves the way this whole decomposition predicts — real at face value, thinning toward nothing once you hold the metabolic and calorie channels fixed. The first meta-analysis to pool the ultra-processed-food-to-dementia link found high (versus low) intake tracked all-cause dementia at RR 1.44 (95% CI 1.09-1.90) — but across a study set so disparate the pooled figure averages almost anything (I2 = 97%), and moderate intake was flat (RR 1.12, 0.96-1.31), so the authors «did not demonstrate a robust dose–response relationship». (Henney et al., 2023) Taken one at a time, every dementia subtype was non-significant — Alzheimer’s 1.08 (0.79-1.48), vascular 2.05 (0.39-10.90), mild cognitive impairment 2.01 (0.75-5.42), dementia excluding MCI 1.24 (0.93-1.65) — so only the pooled all-cause number clears significance. (Henney et al., 2023)
The telling result is what happens on adjustment. Restrict the pool to the studies that adjusted for type 2 diabetes and the association is gone (RR 1.47, 0.97-2.00); restrict to those adjusting for total energy intake and it is gone again (RR 1.26, 0.95-1.67); it also falls apart above a 10,000-participant study-size cutoff, the small-study tell. (Henney et al., 2023) Read against this deliverable’s thesis, that is the answer: the dementia signal is consistent with running through the cardiometabolic disease and surplus calories ultra-processed food carries, not with a processing-specific effect on the brain. (inferred from Henney et al., 2023) It is the same decomposition applied to weight and heart disease, reaching the same place for cognition.
Widen the outcome from a dementia diagnosis to cognition generally and the picture holds. A second review — a narrative synthesis of five observational studies, still a preprint — took in global cognition, individual domains and decline trajectories; three of the five showed an adverse main effect and all five in some subgroup, with no pooled magnitude to rank. (Smith et al., 2026) It is not an independent confirmation: the two reviews share a primary cohort (Li’s UK Biobank), and this one reports its «results largely align with the two other systematic reviews in the field». (Smith et al., 2026)
Its one decision-useful addition is a substitution estimate from that shared cohort — replacing 10% of ultra-processed food by weight with less-processed food tracked a 19% lower dementia risk (HR 0.81, 0.74-0.89), still observational and composition-confounded, but a reminder that the lever is what replaces the food, not the label. (Smith et al., 2026) Feeding trials to a dementia endpoint are ethically impossible, so this arm stays observational by design. The one outcome where a composition-fixed design does exist is intake — and there processing does move the needle.
Processing does move how much you eat — through levers you can name
One randomized trial holds the causal footing, and it names the properties that carry the effect. Hall’s inpatient study fed 20 weight-stable adults an ultra-processed and an unprocessed diet, each matched on presented calories, energy density, macros, sugar, sodium and fibre, and let them eat as much as they wanted. On the processed diet they ate 508 ± 106 kcal/day more and gained 0.9 ± 0.3 kg in two weeks, losing the same on the unprocessed diet (Hall et al., 2019). So processing does independent work on intake at matched composition — but read the endpoint honestly: this is a surrogate (intake and weight, two weeks, n=20), not the mortality or cardiometabolic outcomes the observational category is built on. (inferred from Hall et al., 2019)
The excess was not people liking the processed food more or feeling hungrier. «Both diets were rated similarly on visual analog scales (VASs) with respect to pleasantness and familiarity» (Hall et al., 2019), and hunger, fullness and satisfaction did not separate the diets. The effect ran instead through three measurable properties: the processed food packed 85% higher non-beverage energy density, was eaten faster (the eating-rate gap tracked the intake gap, r = 0.45), and diluted protein enough to explain «at most 50%» of the gap (Hall et al., 2019). This is not NOVA magic — it is energy density, eating rate and protein dilution, each a target a decision can name directly.
One of those levers generalizes well past the single trial. Robinson pooled 22 experiments that manipulated eating rate directly — none of them about processed food — and found a slower rate cut intake: «a slower eating rate was associated with lower energy intake in comparison to a faster eating rate (random-effects SMD: 0.45; 95% CI: 0.25, 0.65; P , 0.0001)» (Robinson et al., 2014). The load-bearing detail is that hunger stayed flat while intake fell — so the channel is mechanical and oral-sensory (duration of taste exposure per unit food), not reward or appetite. It is the same dissociation Hall saw in processed food, reached by a wholly separate design set. (inferred from Robinson et al., 2014)
But the whole randomized base is thin, and that bounds how far the foothold reaches. Aramburu’s systematic review of UPF-reduction trials found the entire literature is four trials, 455 participants, median follow-up 12 weeks — and only Hall directly fed a processed diet. The other three were educational counseling trials, and in two of them processed-food intake was not differentially reduced at all, so their effects belong to the other diet and activity advice bundled in (Aramburu et al., 2024). So the causal claim rests on one small, short, unblindable trial with no confirmatory second feeding study. (inferred from Aramburu et al., 2024)
If the effect runs through nameable properties, the next question is what the “ultra-processed” label adds once those properties are priced — and whether the guideline bodies think it adds anything at all.
Price the components and the label mostly dissolves
Once you price the properties Hall named, most of the category’s predictive power is already accounted for. Two gold bodies make the point as an incremental-validity argument. NNR looked at the same associations and declined to recommend on the category: the classification «does not add to the already existing food classifications and recommendations in NNR2023» (Nordic Council of Ministers, 2023), because a processed diet’s predictive power runs through energy density, added sugar, salt, fat and low fibre — every one a variable NNR already regulates — plus the intake sub-components Hall isolated. Aramburu reaches the same declination from the trial side. If the category predicts through variables you already act on, regulating the category too is double-counting, not new information. (Aramburu et al., 2024; inferred from Nordic Council of Ministers, 2023)
Two specific-exposure nulls show the label dissolving into its parts. Refined versus whole grain is the cleanest: SACN tested refined grains directly and found RR 1.00 (95% CI 0.98, 1.01) for both cardiovascular events and type 2 diabetes (Scientific Advisory Committee on Nutrition, 2015) — a flat null on hard outcomes. The whole-grain benefit that does show up routes through fibre, so fibre content, not the refining step, is the target. Sugar is the same shape: swapping free sugars for other carbohydrate at equal energy moves weight by 0.04 kg (-0.04 to 0.13) — null (World Health Organization, 2015). Sugar’s weight effect is an energy effect; the load-bearing carrier is the sugar-sweetened beverage (per 250 mL/day, type 2 diabetes RR 1.19, 1.13-1.25) (Qin et al., 2020), because a drink adds poorly-compensated liquid calories.
So does “ultra-processed” survive as its own lever on hard outcomes? On the evidence held here, no: the category collapses to its constituents, and the honest instruction is to act on the named components — energy density, eating rate, protein, added sugar, sugary drinks — not the NOVA label. This does not contradict the Hall foothold. Hall’s diet moved intake at matched total composition, but it did so through the identifiable sub-components — the very energy-density and low-fibre properties NNR named as the collinear channel. The boundary does real work on intake, and the rule still holds: replace the label with the measurable property. (inferred from Hall et al., 2019; Nordic Council of Ministers, 2023)
The guidance set disagrees, and that disagreement is itself the finding. Lane’s umbrella review recommends population measures to reduce ultra-processed food; NNR and Aramburu decline to recommend on the category at all. This is not one body being wrong — it is what moved each recommendation. What sits behind the pro-category position is a confounded observational association, not a demonstrated processing mechanism, and where guidance families split on the same evidence, the evidence does not determine the number. The adjudicated position follows from the components: act on the priceable properties, not the label a shopper cannot reliably apply. (Aramburu et al., 2024; inferred from Lane et al., 2024; Nordic Council of Ministers, 2023)
The sharpest test of whether a processing-linked property is its own lever is a construct built to isolate the one Hall could not — the reward property, hyper-palatability.
Hyper-palatability is a real, measurable property — but not yet a proven lever
You can now read whether a food is hyper-palatable off its nutrition label — but no study yet shows that eating fewer such foods changes any health outcome. Fazzino’s contribution is to replace the old vibe words (fast food, sweets) with a data-derived line drawn on nutrient composition. A systematic review of the descriptive literature clustered into three threshold combinations: fat + sodium (> 25% kcal from fat and >= 0.30% sodium by weight), fat + simple sugar (> 20% kcal from fat and > 20% kcal from sugar), and carbohydrate + sodium (> 40% kcal from carbohydrate and >= 0.20% sodium). The load-bearing idea is the combination — every cluster names two ingredients above threshold, none a single nutrient — so the target is nutrient co-occurrence, which single-nutrient guidance (sodium, free sugars, fat) is built to miss. (Fazzino et al., 2019) (inferred from Fazzino et al., 2019)
Read the prevalence as a first-draft line, not a target. Applied to the USDA food database, «In the FNDDS, 62% (4,795/7,757) of foods met HPF criteria» (Fazzino et al., 2019) — but the cut points come from a data-visualization over 75 descriptor items with no advanced statistics, and Fazzino says they «should not be assumed to be fixed or final» (Fazzino et al., 2019). So 62% is where a newly-drawn boundary happens to fall, not a measured constant. The property is preparation, not food identity: the same chicken leg is hyper-palatable cooked in butter and not when grilled — 81% of food types could be prepared to meet the line, and for many the method decided membership. That makes HPF a Layer-3 implementation lever (how a food is cooked), not a food to ban. (Fazzino et al., 2019)
HPF is its own construct — not ultra-processing, not energy density. Ultra-processing (NOVA) is a manufacturing classification; hyper-palatability is a nutrient-composition one, and energy density is kilocalories per gram. They come apart: almost half of HPF items (49%; 2,337/4,809) had low energy density (< 2 kcal/g), so a food can be hyper-palatable and not calorie-dense (vegetables cooked in fat, sweetened yogurt). (Fazzino et al., 2019) Sutton (the same lab — a construct extension, not independent corroboration) measured all three across the US supply 1988-2018: 58-65% of foods were UPF, 55-69% HPF, 37-47% HED, with «moderate to high overlap in foods (40%–70%) across definitions» and «approximately one third of foods… met criteria for all three». (Sutton et al., 2023)
Overlap that high but that incomplete is the point: each construct captures a set the others miss, so a claim proven under one does not transfer to another. The distinctly-hyper-palatable foods are «primarily fresh or whole food items prepared with palatability-enhancing ingredients during cooking» (Sutton et al., 2023) — home-cooked dishes a manufacturing classification cannot see.
On outcomes, HPF is empty. No systematic review links hyper-palatable-food exposure to weight gain, diabetes, or mortality; Fazzino’s own paper is a food-database analysis that calls for its criteria to «be adapted, refined, and tested for their predictive utility». (Fazzino et al., 2019) So HPF earns admission as a specified, measurable exposure — a candidate mechanism — but not a sized lever, and it cannot yet enter the intervention ranking. (inferred from Sutton et al., 2023)
HPF asks what the food is. A sibling literature asks what the eater does — and here the intake evidence has actually located a channel.
Rate, not reward, is where the intake evidence lands
When processing drives over-eating, the demonstrated route is how fast a food is eaten, not how much the eater wants it. Hall’s own ratings point away from reward: the appetite scores a reward story predicts should separate the diets did not — hunger, fullness, satisfaction and capacity-to-eat «were not significant between the diets, suggesting that they did not differ in their subjective appetitive properties», and pleasantness ratings did not separate them either. (Hall et al., 2019) And the eating-rate lever generalizes that dissociation past the single trial — Robinson’s pooled slower-rate effect above cut intake with no hunger penalty — a mechanical, oral-sensory channel, not a reward one.
So the reward story is an open question, not a demonstrated driver. Fazzino proposes hyper-palatability as the mechanism Hall’s trial left unnamed — a reasonable hypothesis. But Hall’s design could not compare hyper-palatable against non-hyper-palatable foods, and the palatability ratings it did collect came out equal. Hall’s appetite-null therefore does not refute the HPF hypothesis, and HPF’s composition line does not explain Hall’s rate finding: the two answer different questions (does composition drive intake vs does processing drive intake at matched composition), so neither settles the other. The honest position is that eating rate is evidenced and reward is unproven — not that reward is ruled out. (Fazzino et al., 2019; inferred from Hall et al., 2019)
Food addiction — a contested construct with no outcome behind it
Food addiction is a different object again: a behavioural construct (loss-of-control eating, measured by the self-report Yale Food Addiction Scale mapping substance-dependence criteria onto food), not a composition definition — so it does not transfer to HPF or UPF either. Schulte found three food attributes predict a high addictive-like-eating score — processing, fat, and glycemic load, all large effects — and named the implicated foods those «designed to be particularly rewarding through the addition of fat and/or refined carbohydrates». (Schulte et al., 2015)
A gold systematic review (Gordon) «generally support[s] the validity of food addiction as a diagnostic construct» — but read the bar it clears: the support is vote-counted (some criteria rest on a single study, brain-reward on 21), heavily animal, and the review flags its own confirmation-bias risk while opening by calling the subject «a highly controversial subject». (Gordon et al., 2018) Schulte and Gordon share the YFAS lineage — Gordon’s evidence base literally includes Schulte — so their agreement is shared-lineage coherence, not independent convergence. (inferred from Gordon et al., 2018)
A prevalence review (Pursey) pooled 25 studies to a «weighted mean prevalence of FA diagnosis in adult population samples… 19.9%» (Pursey et al., 2014) — but the pool was «predominantly comprised of overweight/obese females recruited from clinical settings», so «the prevalence… and the average symptom scores are likely higher compared to a nationally representative general population sample». (Pursey et al., 2014) Schulte’s own community rates (6.7% undergraduates, 10.2% online adults) sit well below it, so 19.9% is a clinically-enriched figure, not a population rate — and how common a self-report symptom count is says nothing about whether the construct is valid or harms anyone.
On patient-important outcomes the construct is empty: no review links addictive-like eating to weight gain, diabetes, or mortality as a pathway distinct from ordinary over-consumption, so no decision turns on it beyond the ordinary reasoning already covered. (inferred from Pursey et al., 2014; Schulte et al., 2015)
The remaining aspects — additives and contaminants — are where mechanism runs ahead of human outcome data.
Additives and contaminants: mechanism outruns outcomes
Every additive on this list has a plausible mechanism and almost none has a hard-outcome result. That gap is the whole point: a named biochemical pathway is a reason to keep watching a compound, not evidence that it harms the people who eat it. Judge each additive on its own evidence, one compound at a time — “it is an additive, therefore it is harmful” is the fallacy the mechanism stories keep feeding.
Non-sugar sweeteners: the comparator decides the effect. WHO’s 2023 guideline «suggests that non-sugar sweeteners not be used as a means of achieving weight control or reducing the risk of noncommunicable diseases (conditional recommendation)» — a suggestion against on low-certainty evidence, not a finding of harm. (World Health Organization, 2023) The pooled trial benefit of -0.71 kg (95% CI -1.13 to -0.28, 29 RCTs, n=2433) exists only when a sweetener displaces sugar; against water it disappears, and in the trials that actually asked habitual sugar consumers to switch it attenuates to -0.61 kg (95% CI -1.28 to 0.06), non-significant. (World Health Organization, 2023) So the sweetener does no work the sugar it replaces was not already doing.
No demonstrated population harm — and a genuine open question, not a resolved one. Short-term trials show flat cardiometabolic biomarkers; long-term cohorts show opposite-signed associations with hard disease — incident obesity HR 1.76 (1.25 to 2.49), type 2 diabetes HR 1.23 (1.14 to 1.32), all-cause mortality HR 1.12 (1.05 to 1.19), all at low-to-very-low certainty, with overall cancer null (HR ~1.02). (World Health Organization, 2023) WHO ran the reverse-causation check and the association weakened but survived — so the long-term signal is unexplained, cannot be dismissed as reverse causation, and is not established as causal. The honest verdict is a suggestion against use for weight control, not a harm claim.
Aspartame is the textbook hazard-vs-risk case, and it defuses the 2023 cancer scare. IARC’s 2023 review placed aspartame in Group 2B — “possibly carcinogenic,” the weakest positive tier, on limited human evidence that could not rule out chance, bias or confounding — which flags only whether the compound can cause cancer under any condition, not what happens at consumed doses. (Riboli et al., 2023) JECFA, reviewing the same evidence the same week, reaffirmed the acceptable daily intake of 0-40 mg/kg/day and found realistic intake (adults ~5 mg/kg/day at the mean, ~12 for high consumers) does not approach it, so a weak can-it-ever-in-principle hazard signal is fully consistent with safe at the doses people actually eat — the risk at consumed intake is not demonstrated. (Joint FAO WHO Expert Committee on Food Additives, 2023) -> Non-Sugar Sweeteners
Emulsifiers, titanium dioxide (E171) and nitrites: signal without outcomes, and it must be read per compound. Each carries a directional mechanistic case, none carries demonstrated human hard-outcome harm, and lumping them under “additives” erases the distinctions that matter. Nitrite is the cleanest example of why the lump misleads: “cut the nitrite” names three different exposures with different signs. Free dietary nitrate and nitrite — over 80% vegetable-borne — show no colorectal signal (colon nitrite OR 1.02, 0.92 to 1.11; rectal 1.09, 0.79 to 1.39), so cutting it by eating fewer vegetables would be net-harmful. (Said Abasse et al., 2022) The candidate colorectal channel is the in-matrix curing nitrite that nitrosylates heme in processed meat — a different molecule in a different place, addressed in the next section.
Dietary AGEs and low-heat cooking: the surrogates barely move and the outcomes that matter do not. Baye’s meta-analysis of 17 RCTs (560 participants) found low-AGE diets «decreased insulin resistance (mean difference [MD] -1.3, 95% CI -2.3, -0.2), total cholesterol (MD -8.5 mg/dl, 95% CI -9.5, -7.4) and low-density lipoprotein (MD -2.4 mg/dl, 95% CI -3.4, -1.3)» — an LDL drop of about -0.06 mmol/L, trivially small. (Baye et al., 2017) The headline parameters were flat: «There were no changes in weight, fasting glucose, 2-h glucose and insulin, haemoglobin A1c, high-density lipoprotein or blood pressure.» (Baye et al., 2017) Every endpoint is a surrogate; no event or mortality trial of AGE reduction exists, so the path from fewer dietary AGEs to a patient-important outcome is unevidenced.
“Thin” means insufficient, not safe — and the two failure directions are symmetric. Do not let the additive nulls read as an all-clear, and do not let the mechanistic pathways read as demonstrated harm; both over-read the evidence, in opposite directions. The wiki’s job at layer 2 is to name the state — insufficient evidence, per compound — and the precautionary weighting of an unresolved-but-plausible signal belongs to the person at layer 3, not to the appraisal. Which raises the question the additives cannot answer: is there a whole-food category that carries an evidenced hard-outcome effect? One does.
Processed meat is the one category with an evidenced harm
Of every category in this document, processed meat is the one with a real, if modest, hard-outcome signal. Processed meat raises colorectal-cancer risk by RR 1.16 (95% CI 1.08 to 1.26) per 50 g/day — significant and consistent across cohorts. (World Cancer Research Fund International, 2018) Red meat is weaker and does not clear the bar: RR 1.12 (95% CI 1.00 to 1.25) per 100 g/day, a pooled estimate whose lower bound touches the null, so the association is not statistically significant in WCRF’s own words. (World Cancer Research Fund International, 2018) The evidenced harm attaches to the processed category specifically, not to red meat in general.
The harm has a nameable sub-component, which is what makes it more than a category average. Heme iron — present in red meat «10-fold higher than that of white meat» — acts as a catalyst for endogenous N-nitroso formation and lipid peroxidation, the proposed reason red meat carries colorectal risk and white meat does not. (Bastide et al., 2011) Highest-versus-lowest heme intake carries a colon-cancer RR of 1.18 (95% CI 1.06 to 1.32), though this is a categorical contrast, not a dose-response curve, and partly a red-meat proxy. (Bastide et al., 2011) In processed meat the curing nitrite nitrosylates that heme, coupling the two channels — which is why the additive section’s “free dietary nitrite” null does not rebut the meat-matrix mechanism: they are different exposures.
The absolute effect is small, and what the numbers warrant is a separate, contested question. Re-pooled into absolutes for a realistic 3-servings/week reduction, processed meat maps to roughly 1 to 8 fewer cancers and 1 to 12 fewer cardiometabolic events per 1000 people over a lifetime, at low certainty. (Johnston et al., 2019) WCRF reads this evidence as warranting a limit — «there is no level of intake that can confidently be associated with a lack of risk of colorectal cancer» — while NutriRECS reads the same cohorts as too low-certainty to change what adults do. (World Cancer Research Fund & American Institute for Cancer Research, 2018) That clash is a grading-and-standpoint disagreement about a shared evidence base, not a dispute about the effect size, and it is worked out in full on Should Adults Reduce Red and Processed Meat. So where does the whole decomposition leave a person deciding what to eat, measured against the big rocks?
What to do — act on the components, against the big rocks
Stop shopping for the NOVA label and buy the properties instead. The evidence licenses a short, priceable list: reduce energy density (especially in food, not drink), slow the eating rate or choose harder textures, keep protein adequate, cut added sugar and sugary drinks, and limit processed meat. Each of these is something a person can see, measure, and act on at the shelf — unlike “avoid ultra-processed food,” a category boundary a shopper cannot reliably apply and which, on hard outcomes, collapses into these very components. These are not additional levers layered on top of “reduce UPF”; they are what “reduce UPF” reduces to once the confounding is stripped out.
Judge every one of these against the realistic substitute, not against an ideal. The effect of removing a food depends entirely on what replaces it: a sweetener earns its small benefit only when it displaces sugar and none when it displaces water; cutting processed meat helps or not depending on what fills the plate instead. State the swap, price it against the alternative the person would actually reach for, and the decision becomes tractable in a way “eat less processed food” never is.
None of this is a big rock, and pretending otherwise misdirects attention. The ranking of levers is dominated by a handful of large exposures — smoking, obesity, near-total inactivity — and where one of those stands unaddressed, no amount of processing precision changes what a person should do next. The population evidence agrees from the outside: red meat, processed meat, trans fat and sugar-sweetened beverages sit «towards the bottom» of the attributable-burden ranking, while the boring under-eaten staples sit at the top — a worked instance of attention running inversely to effect size. (Afshin et al., 2019) The volume of discourse around ultra-processing is a fact about the field, not about the size of the lever.
Two honest limits close the account. First, the outcome many people care about most — the shape of decline, full function then a rapid end rather than a slow multi-decade loss — is unmeasured here; the meat and intake trials count events and surrogates, not trajectories, so the wiki names this as a gap rather than letting a survival endpoint stand in for a life well-shaped. Second, for a person already lean, active, non-smoking and sleeping well, the big rocks are already pulled and the processing question is a small, honest lever — and saying so is itself a result. It licenses that person to stop optimizing the label and reclaim the attention for something that would move an outcome they care about.
Evidence box
Question ’What does the evidence show about processed and ultra-processed food”s effect on each patient-important outcome — in which direction, how large, for whom, how certain — once “processed” is decomposed into the distinct aspects (matrix disruption, additives, refining/substrate change, energy density, hyper-palatability as an engineered reward property distinct from ultra-processing, processing contaminants) and the distinct categories that differ in evidence? How does the effect vary by aspect and by category, and how large is any effect that survives the observational caveats relative to the big rocks?‘ Evidence included 26 sources — 14 gold, 9 high, 3 moderate Overall certainty Low (see Rating Certainty of Evidence) Source-selection note 3 source(s) below the gold evidence bar feed this page: Fazzino (cohort, moderate); Sutton (cohort, moderate); Schulte (cohort, moderate). Each labelled by tier; none load-bearing for the core claims. Last updated 2026-09-04 · Independently reviewed: No · Full edit history