The Controlled-Feeding Trial: Why People Ate More

Most of what we know about ultra-processed food comes from watching large populations — consistent, suggestive, and vulnerable to confounding. In 2019, one team tested it causally: check twenty adults into a research hospital for a month, feed them two menus matched for nutrients, and let them eat as much as they want. They ate roughly five hundred more calories per day on the ultra-processed menu and gained weight while losing it on the unprocessed one. This page takes that experiment apart — what makes the design strong, why the number matters, and exactly where its reach ends.

The Study in One Paragraph

Kevin Hall's laboratory at the National Institute of Diabetes and Digestive and Kidney Diseases ran the question through the cleanest design nutrition research has: twenty weight-stable adults, average age about thirty-one, BMI around 27, admitted to the NIH Clinical Center for four weeks. For two weeks each, in randomized order, every meal came from one of two dietitian-designed menus matched for presented calories, energy density, macronutrients, sugar, sodium, and fiber. One menu drew roughly four-fifths of its energy from ultra-processed items — the NOVA category defined on this topic's first page — and the other essentially none. Meals and snacks arrived at double each person's estimated maintenance needs on rotating seven-day menus; participants ate as much or as little as they desired. Nobody cooked for themselves, nobody recalled portions from memory, and nobody went home at night.

Why the Design Earns Its Reputation

Nutrition trials usually leak — people forget what they ate, sneak food, or drop out. This one plugged the usual holes:

• 🔁 Crossover, self-controlled: each participant ate both diets in random order, serving as their own comparison — genotype, habit, and baseline appetite largely cancel out.
• 🏥 True inpatient control: four weeks on a metabolic ward — every calorie served was weighed; none came from a vending machine.
• 🧪 Matched menus: the diets were matched for calories, energy density, macronutrients, sugar, sodium, and fiber — "ultra-processed" was the headline difference, not a nutrient proxy.
• 🍽️ Ad libitum intake: the outcome was what people freely chose to eat — the pathway through which ultra-processed food is hypothesized to act.
• ⚖️ Objective endpoints: intake was measured by weighing what came back; weight came from ward scales, not questionnaires.

One classic feature is missing: blinding. Participants knew which diet they were on — you cannot disguise a salad as a hot dog — leaving room for expectation effects, though the intake gap is large enough to make a pure expectation story hard to sustain.

What Happened: About Five Hundred Calories a Day

On the ultra-processed arm, participants consumed 508 ± 106 more calories per day than on the unprocessed arm (p = 0.0001). Nearly all of the surplus came from carbohydrate (about 280 kcal/day) and fat (about 230 kcal/day); protein was statistically unchanged. Weight followed energy: participants gained an average of 0.9 kg across the two ultra-processed weeks and lost 0.9 kg on the unprocessed ones; individual weight change tracked intake difference tightly (r = 0.8). That is the signature of a real energetic mechanism — appearing without anyone rating the ultra-processed menu as tastier or more familiar.

Why They Ate More: The Leading Explanations

The trial established that the overeating happened, not why — but secondary measurements point at a short list of suspects the mechanisms page develops in full:

• ⏩ Eating speed: participants ate significantly faster on the ultra-processed diet — in both kilocalories and grams per minute — and those who sped up more tended to overeat more (r ≈ 0.45). Soft, ready-to-swallow food delivers calories before fullness signals catch up.
• 🥩 Protein dilution: absolute protein intake was nearly identical on both diets even as protein's share of energy fell — the pattern protein-leverage hypotheses predict. Consistent with the data; not isolated by it.
• 🥤 Beverage energy density: overall energy density was matched, but liquid calories behave differently from solids; the authors flag non-beverage energy density as imperfectly matched.
• 😋 Hyper-palatability, examined and not convicted: participants rated the diets equally pleasant and familiar, and appetite ratings did not differ — so the simple "tastier food" story got no support, though a questionnaire is a blunt instrument for supernormal appeal.
• 🧪 Additives: emulsifiers, sweeteners, and texture modifiers traveled with the menu but were never tested as isolates; no single ingredient class can be convicted here.

What the Trial Does Not Establish

The result is the strongest causal evidence in this literature — and it is exactly one month long. Both facts matter:

• 📅 No long-run outcomes: two weeks per arm cannot show whether the intake excess persists, fades with adaptation, or accumulates into disease — those endpoints belong to the cohorts the outcome-evidence page prices honestly.
• 👥 Twenty specific adults: midlife, weight-stable, BMI around 27, healthy enough for admission; no children, no diabetes, no medications. Whether the effect size generalizes is an assumption, not a finding.
• 🧩 Whole diets, not properties: the menus differed on many axes at once — texture, speed, protein share, beverage density, additives. The trial shows the package drives overeating; it cannot pin the effect on "ultra-processedness" as a single thing, nor promise every NOVA-4 item behaves like this menu.
• 🏠 Ward, not world: free-living eaters face cost, access, and habit pressures the ward removed — and they cut the wrong way: the study priced the ultra-processed menu at about $106 per week per 2,000 kcal/day versus $151 for the unprocessed one.
• ✍️ The authors' own caveats: the paper's own discussion names imperfect protein-percentage and non-beverage energy-density matching as limitations future trials should tighten.

Using the Finding Without Overusing It

• 📉 Move the share, not the purity: the actionable translation is dietary share — shifting from half of calories toward a quarter, as the parent topic frames it; zero is not a target this evidence supports.
• 🐢 Slow the meal: eating speed was the trial's most concrete behavioral signal. Food that needs chewing buys time for fullness signals — plate it, pause between bites, set the fork down.
• 🥗 Anchor meals on protein and fiber: protein intake self-regulated here even as calories drifted, and fiber-rich meals slow energy delivery — mechanics the fiber page and the fats-and-carbs page cover; fiber was matched in this trial, but remains a practical anchor.
• 🍬 Watch the refined-carbohydrate vector: most of the surplus was carbohydrate and fat — the same delivery routes fructose's special case explains.
• 🛒 Keep defensible engineered foods: some fortified and processed items earn their place — the edge-case test sorts them item by item, and a chronic surplus landing where the liver gets fat is the downstream stake.

Questions, Answered Briefly

• 😋 "Didn't the processed menus just taste better?" Participants rated pleasantness and familiarity equal, and appetite ratings did not differ — the overeating happened without measurable palatability differences, though subtle appeal beyond questionnaires can't be ruled out.
• 🧾 "Is this a license to call UPF the obesity cause?" It establishes a causal effect on intake over two weeks in twenty adults. Population obesity has many drivers; this trial is the mechanism bridge, not the whole ledger.
• 💸 "Was the unprocessed menu realistic on cost?" It ran about 40% more per week at supermarket prices — a real structural barrier; frozen produce, dried beans, and batch cooking narrow the gap.

Related Topics