Researchers weighed 2,733 meals to the nearest 0.1 gram. Three food characteristics predicted 76% of how much people ate — and none of them appear on any nutrition label.
Scientists predicted 76% of how much people ate using three things no calorie app tracks: calories per gram, grams per minute, and percentage of hyper-palatable food. The tracking dashboard was not wrong. It was incomplete.
Researchers at the National Institutes of Health weighed 2,733 meals to the nearest tenth of a gram. Then they built a model.
Using three traits of each meal — none of which appear on any food label, in any calorie app, or in any diet plan — the model predicted how much each person ate within roughly 170 calories.
The match between predicted and actual intake: 0.76 out of 1.0. Across four diets, 35 adults eating in a metabolic research ward, and two weeks per diet in random order. Three variables explained more about meal intake than everything most people track. All three were invisible to every tracking tool those people could have used.
The three strongest predictors of how much you eat at a meal are invisible to every tracking tool — and one of them, protein, pulled intake in the opposite direction from what most people expect.
- The food characteristic with the strongest effect on how much you eat is one no nutrition label shows: how many calories are packed into each gram of food.
- In two of four diets tested, higher protein percentage was linked to eating more at each meal, not less — contradicting the most common piece of nutrition advice.
- After a large meal, your body normally corrects by eating 26–36% less next time. In the ultraprocessed diet, that correction was essentially zero.
- The two strongest drivers of intake partially share the same mechanism — meaning a food that scores high on both doesn't hit you with two full-strength effects.
The Variable That Matters Most
The strongest of the three predictors was not what most people would guess.
It was energy density — how many calories are packed into each gram of food. Across all 2,733 meals and all four diets, energy density had the largest effect of any single variable on meal intake. More than double the next strongest predictor. The probability of this being noise: less than 1 in 10,000.
Two meals can look identical on a plate. Same volume. Same weight. But if one carries 1.2 calories per gram and the other carries 2.4, the denser meal drives people to eat more. Not because the person chose to. Density pulls eating up across every diet tested.
Energy density is not listed on any nutrition label. No calorie app calculates it for your meal. The strongest single predictor of how much you eat is a number you have never been asked to look at.
The Food Category That Doesn't Play Fair
The second-strongest predictor was the percentage of each meal that qualified as hyper-palatable food — foods crossing specific nutrient-pair thresholds that create a stronger pull than any single ingredient.
Fat combined with sugar. Fat combined with sodium. Carbohydrates combined with sodium. When any of these pairs crosses a set threshold, the food meets the criteria. Pizza, ice cream, bacon, crackers — but also granola bars, flavored yogurt, and many foods marketed as healthy.
Roughly half as strong as energy density, but significant across all four diets. Even after accounting for energy density and eating rate, hyper-palatable food independently drove people to eat more at each meal.
How common are these foods? The researcher behind this study had applied the definition to the entire US food database [1]. 62% of foods met the criteria.
Fat-sodium pairs made up 70%. Fat-sugar pairs: 25%. Carb-sodium: 16%.
And of all foods marketed as low-sugar, reduced-fat, or no-sodium — foods people buy specifically to eat healthier — 49% still qualified as hyper-palatable [1].
The Clock Nobody Watches
The third predictor was eating rate — grams consumed per minute.
Significant in every single dietary pattern tested. The effect was weakest in the ultraprocessed diet and nearly four times stronger in the low-carbohydrate diet.
A meta-analysis of 22 experimental studies confirmed the same direction independently [2]. Slower eating reliably reduced intake across four different methods of manipulating speed — verbal instructions, computerized feedback, food texture changes, and food delivery method. The effect held regardless of approach [2].
Eight minutes at your desk. Lunch gone. Hungry again in ninety minutes. The calories on the tracking app were correct. The speed was the untracked variable doing the work.
The two strongest drivers — energy density and hyper-palatable food percentage — partially overlap in their effects. A food that scores high on both doesn't hit you with two full-strength punches: the combined impact is real, but less than the sum of the individual effects.
The Lever That Pulls Both Ways
Here is where the data does something unexpected.
Protein percentage — the most commonly recommended dietary lever for controlling appetite — was linked to eating more at meals. Not less. More. In two of the four diets.
In the ultraprocessed diet, higher protein share was linked to eating substantially more, not less. The same pattern appeared in the unprocessed diet. Both with a probability of less than 1 in 10,000 of being noise.
In the low-fat and low-carbohydrate diets? No measurable effect. Functionally zero.
If the single most repeated piece of nutrition advice is "eat more protein to feel full," this finding creates a specific kind of vertigo. Within these meals, the protein lever pulled consumption in the opposite direction.
But the paradox has a partial resolution. The study also tracked what happened between meals. In the ultraprocessed diet — and only there — higher protein at one meal reduced how much people ate at the next one. The lever that pushed consumption up within a meal pulled it back down between meals.
A split most people have never thought about: within-meal intake and between-meal satiety are different mechanisms. Higher protein in a meal did not stop people from eating more of that meal. But in the ultraprocessed diet, it reduced what they ate at the next one.
The researchers flagged the gap themselves. In their words: "Perhaps higher protein decreases energy intake only when eating diets high in ultraprocessed foods, or perhaps the protein effects require longer to manifest."
The Diet Where the Brake Doesn't Work
Your body has a built-in brake. Eat a large meal and you tend to eat less at the next one. The system self-corrects.
In three of the four diets, this worked. Previous meal energy intake reduced subsequent meal intake by 26% to 36% — low-carbohydrate (36%), low-fat (30%), and unprocessed (26%). All highly significant.
In the ultraprocessed diet: 4%. No different from zero.
A big ultraprocessed lunch did not reduce ultraprocessed dinner. The brake that engages in every other dietary pattern — the system that claws back a quarter to a third of the overshoot — failed to activate.
And people on the ultraprocessed diet also ate faster in the parent study [3]. Faster eating and broken brakes. Two independent failures stacking.
Every time you have felt unable to stop eating processed food and blamed your discipline — the data suggests the brake was disengaged. Not by you. By the food.
The brake failure compounds over time. Dieting itself changes the hormonal landscape: a year after weight loss, GLP-1 had decreased, ghrelin was still elevated, and the fullness signal PYY had dropped further than at week 10. GLP-1 is the same hormone this study found linked to eating rate in all four diets. A post-dieter eating ultraprocessed food faces two independent systems working against them: brakes that don't engage and a hormonal committee that already voted 8-to-1 for regain.
The Brake That Didn't Engage
Why People Ate 500 Extra Calories a Day
This study did not appear out of nowhere. Fazzino's team reanalyzed data from the first trial to prove that ultraprocessed food causes weight gain [3].
In that 2019 study, Kevin Hall and colleagues at the NIH found that people ate about 500 calories more per day on an ultraprocessed diet. The diets were matched on total calories offered, sugars, fiber, fat, and carbs. Participants gained 0.9 kg in two weeks [3].
That study proved the effect. It left the mechanism open.
Fazzino's model fills the gap. Energy density on its own explained 45.1% of the intake difference between ultraprocessed and unprocessed diets. Hyper-palatable food percentage alone explained 41.9%. Each tested on its own — the numbers cannot be added, but each one accounts for a large share of the gap.
The 500-calorie difference was not explained by the macros. It was explained by three food characteristics the macros do not capture.
In three diets, your body auto-corrected after a big meal — eating 26–36% less next time. In the ultraprocessed diet? Four percent. Indistinguishable from zero. The brake was disengaged.
The Researcher Who Connected All Three Dots
Tera Fazzino works at the Cofrin Logan Center for Addiction Research and Treatment at the University of Kansas. She studies food through an addiction lens, not just a nutrition one.
In 2019, she created the formal definition of hyper-palatable food — the thresholds this study uses [1]. In 2023, she proved those foods independently drive overeating across four dietary patterns. And in 2024, she published a finding that connects the overeating data to something most nutrition pages would never think to look for [4].
US tobacco companies, during their decades of investment in the American food industry, pushed hyper-palatable foods into the food supply. Tobacco-owned food brands were 29% more likely to be fat-sodium hyper-palatable and 80% more likely to be carbohydrate-sodium hyper-palatable than brands that were never tobacco-owned [4].
One scientist. Three papers. Three decades of dots connected: defined the problem, measured it in 2,733 meals, and traced its origin to an industry that spent decades engineering addictive products in a different category.
The tobacco companies divested from food in the early 2000s. The hyper-palatable food supply they helped shape has not changed.
What This Study Cannot Prove
The researchers were explicit: "The effects observed indicate associations that are not necessarily causal." This was a secondary analysis of meals already eaten in two prior studies.
Thirty-five adults ate alone in inpatient rooms at the NIH — a setting that removes the social, emotional, and environmental factors that shape eating in real life. The model's 76% correlation is strong. The 170-calorie mean error is roughly 23% of the average meal — meaningful but not precise.
The protein finding carries the researchers' own doubt, and those percentages describe links, not proven causes. And a new trial is underway (NCT05290064) to test whether changing these food traits actually changes how much people eat.
That honesty is what makes the data worth taking seriously. A page that hides its caveats is selling something. This study is showing you what 2,733 meals revealed — including the parts that remain uncertain.
Three food characteristics predicted 76% of meal intake across four diets. None of them appear on any nutrition label. The tracking dashboard — calories, macros, portion sizes — was not wrong. It was incomplete.
Three new variables. All measurable. None requiring a new app or a new diet. Just a wider lens on the plate that is already in front of you.
And if energy density is the strongest of the three — the one variable with the largest effect in every dietary pattern tested — the natural next question is whether you can eat more food by choosing meals with lower energy density. That question has its own meta-analysis.
The study's three predictors are not abstract research variables. Energy density is visible on every plate — a pile of vegetables versus the same-sized portion of cheese tells the story. Eating speed is something you can notice in real time. And hyper-palatable food has a practical test: does this food combine fat with sugar, fat with salt, or carbs with salt in proportions that make it hard to stop eating?
No new tools required. Just three questions the research says matter more than the ones most people currently ask.
What other research found
What this means for you
The data paints a specific picture for this dietary pattern. Your body's meal-to-meal correction system was effectively turned off — compensation was statistically indistinguishable from zero, compared to 26–36% in the other three diets.
On top of that, people eating ultraprocessed food in the parent study ate faster. Two independent failures stacking: faster eating within each meal and no compensation between meals. These are different mechanisms operating at different timescales, both specific to ultraprocessed food.
The three drivers predicted intake even in the unprocessed diet — this is not exclusively a processed food problem. Energy density and hyper-palatable food percentage still drove how much people ate, just with smaller effects.
The protein paradox applied here too: higher protein percentage at each meal was linked to eating more within that meal. And the compensation mechanism worked, but only corrected about a quarter of the overshoot — not all of it.
Eating speed had the strongest effect in the low-carb diet — nearly four times the effect seen in the ultraprocessed diet. The protein paradox did not apply here (the protein effect was functionally zero).
The compensation mechanism worked best in this pattern, correcting about 36% of the previous meal's excess. An interesting tension: low-carb foods tend to require more chewing, which may naturally slow eating — but liquid fats and keto shakes bypass this protection entirely.
Before you change anything
Thirty-five adults aged 18 to 50, weight-stable, with an average BMI around 27–28 (overweight range). All lived at the NIH research facility for the duration of each diet period. Roughly equal numbers of men and women.
The study did not include children, adolescents, adults over 50, athletes, or anyone with an eating disorder or metabolic condition. And the meals were prepared and served — nobody chose their own food.
This was a secondary analysis — the original studies were designed to compare diets, not to test whether energy density, eating speed, and hyper-palatability predict intake. The researchers reused data from meals that had already been eaten.
Every meal was eaten alone in an inpatient room. Social eating — which reliably increases how much people consume — was completely absent. And the menus were controlled: the model predicts intake of food that was offered, not whether people would choose this food in real life.
The protein finding carries the researchers' own uncertainty about its mechanism and the conditions under which it applies.
Stronger than a survey but weaker than a purpose-built trial. The data came from two randomized crossover studies where every meal was weighed to the nearest tenth of a gram — no dietary recall, no self-reporting. That level of measurement control is rare.
But the prediction model was not what these studies were designed to test. It was built after the fact from data collected for a different purpose. A new trial (NCT05290064) is underway to test whether changing these food characteristics actually changes how much people eat — closing the gap between association and cause.
Energy density had the largest effect in every diet tested. The natural next question: if density matters most, can you eat more food by choosing meals with lower energy density?
That question has its own meta-analysis — 31 studies spanning decades of controlled feeding research. The answer involves a daily calorie difference that most people find hard to believe.
What This Study Found
All findings from this paper, in plain language.
- The single strongest predictor of how much people ate was how many calories were packed into each gram of food — not total calories, not portion size.
- Foods crossing specific combinations of fat with sugar, fat with salt, or carbs with salt independently drove people to eat more, even after accounting for calorie density and eating speed.
- Faster eating consistently predicted larger meals across all four dietary patterns, with the effect nearly four times stronger on a low-carb diet than on an ultraprocessed one.
- Higher protein percentage was linked to eating more within each meal in two diets — the opposite direction from what most nutrition advice predicts.
- The effects of calorie density and hyper-palatable food partially overlapped, meaning a food high in both doesn't drive intake as much as the individual numbers would suggest.
- Together, the four meal characteristics predicted intake with 76% accuracy — better than any single food property could achieve alone.
- Energy density individually accounted for about 45% of the intake gap between ultraprocessed and unprocessed diets, and hyper-palatable food accounted for about 42%, each tested separately.
- In three diets, eating a big meal reduced the next meal by 26–36%. In the ultraprocessed diet, compensation was essentially zero.
- The eating speed effect was strongest in the low-carbohydrate diet, where it was nearly four times larger than in the ultraprocessed diet.
- Beverages contributed only about 3% of meal energy and did not change the results when included.
- Previous-meal protein reduced next-meal intake in the ultraprocessed diet but increased it in the low-fat and low-carb diets — the effect depended on what else was being eaten.