“Reliable” can mean three different things: Do two people assign the same group? Does the group describe what it claims to measure? Does it predict something useful for health? Combining these questions turns two apparently contradictory studies into a battle of beliefs. Separating them supports a less dramatic but more robust conclusion: Nova works better as an explicit population-level framework than as an automatic verdict printed on every package.
Compare what Nutri-Score and Nova actually measure
1. Reproducibility, validity, and usefulness are different questions
Reproducibility asks whether raters place the same food in the same group. Validity asks whether that group corresponds to the stated concept—formulation, purpose, and processing—or to an independent reference framework. Utility asks whether the classification actually helps research, policy, or food choices.
High agreement does not prove that a boundary is biologically correct. Coders can apply an imperfect rule consistently. Conversely, a somewhat fuzzy category may still detect a major population-level shift in the food supply if classification errors do not erase the overall signal.
Predicting health is another step. An association between high Group 4 intake and disease does not show that every product was classified perfectly or that processing alone explains the relationship. Nutritional profile, amount, food matrix, social habits, and other factors may act together.
2. Nova has a clear concept, but its boundaries use heterogeneous criteria
The definition published by Nova’s developers covers the nature, extent, and purpose of processing. Group 4 particularly describes industrial formulations, substances rarely used in home cooking, and additives intended to alter sensory properties. This concept adds a dimension that nutrient data alone do not capture. Monteiro et al., 2016 — definition of the four Nova groups, Monteiro et al., 2019 — Nova 4 identification criteria
Problems emerge at the boundaries: How many ingredients constitute a formulation? Which additive serves only as a preservative? What counts as “rarely” used in cooking? How can a coder know about a process absent from the label? Examples help, but a non-exhaustive list does not always produce a clear decision rule.
The framework also combines the food itself, how it was manufactured, and its presumed purpose. Packaged bread, yogurt, a plant-based preparation, or a generic mixed dish may change groups depending on ingredients and origin. When a dietary survey does not retain those details, researchers must reconstruct or assume part of the exposure.
3. Braesco et al. found low agreement among specialists in 2022
Braesco and colleagues asked French specialists to classify 120 commercial products with ingredient lists and 111 generic foods. Fleiss’s κ was 0.32 for products and 0.34 for generic foods. Approximately one-quarter of the foods received particularly inconsistent assignments. Braesco et al., 2022 — operational robustness of Nova
The result documents a real weakness when people apply the provided criteria to the available information. Yogurts, certain mixed dishes, and products with unknown home or industrial origins illustrated the uncertainty. The study did not measure best-case performance after standardized training, however, and it lacked an independent gold standard for declaring one group “correct.”
Funding and interests should be visible: less than 5% of the cost of the survey tool and statistical work came from the French Fund for Food and Health, which was partly supported by industry at the time. The funder was reported to have no scientific role. Several authors stated that their activities could receive food-industry funding or compensation, and two worked for MS-Nutrition. These disclosures warrant transparency, not automatic rejection of the data. French version archived in HAL — disclosed funding and interests
4. Sneed et al. found much higher agreement after training
In a study published in 2023, trained and certified coder pairs classified 3,099 unique foods from 24-hour dietary recalls. Raw agreement reached 88.3%, with a Cohen’s κ of 0.75. The team then adjudicated discrepancies. Sneed et al., 2023 — reliability of Nova coding after training
This finding does not automatically refute Braesco. The foods, information, training, and procedures differed. Sneed evaluated a structured coding method in nutrition software using dietary recalls from children; Braesco tested agreement across a large group of professionals classifying commercial products and generic foods.
Together, the studies offer an operational lesson: Nova should not be applied by intuition. Detailed rules, training, double coding, and adjudication greatly improve consistency. Research teams can build that infrastructure; consumers facing a partial ingredient list usually cannot.
5. A 2026 systematic review confirms that performance depends heavily on protocol
A systematic review published in 2026 included 30 studies. Reported test-retest reproducibility coefficients ranged from approximately 0.46–0.94; validity correlations, from 0.47–0.72; and agreement between classification systems, from 0.36–0.84 depending on the statistic used. Fifteen studies had a low risk of bias and 15 had an unclear risk. 2026 systematic review — reliability and validity of processing classifications
These ranges are not an “average accuracy” that can be assigned to Nova. They combine different instruments, populations, and comparisons. Instead, they show that measurement can be acceptable under some conditions and fragile under others, particularly when recipe or preparation details are missing.
Validity also depends on the reference framework. Comparing Nova with another classification does not reveal which one measures the “true” processing level when the systems define the concept differently. Progress requires more detailed product data, explicit criteria, and sensitivity analyses showing whether conclusions survive ambiguous cases.
6. ANSES acknowledges a health signal while calling for better characterization
In an opinion signed in November 2024 and published in January 2025, ANSES noted the lack of a consensus definition. It considered Nova more focused on formulation than on processing and found Group 4 criteria insufficiently precise, non-exhaustive, and without a hazard threshold. Context-dependent examples can lead to subjective classification. ANSES — full opinion on so-called ultra-processed foods
The agency did not conclude that every relationship with health is imaginary. It acknowledged epidemiological associations but considered the overall weight of evidence low for several relationships and recommended identifying contributing factors separately: formulation, processes, newly formed substances, additives, food matrix, and other exposures.
“Low” here describes certainty in the evidence base, not the size of a problem or the absence of an association. The reasonable conclusion is neither “Nova is scientifically perfect” nor “Nova has been invalidated.” The classification made an important issue visible, but its operational criteria and possible mechanisms still require refinement.
7. An association with health does not validate every classification boundary
Cohort studies classify intake and then compare groups of people across several years. Classification error can weaken, amplify, or shift an association depending on its structure. A signal repeated across populations strengthens interest in the question without directly testing the accuracy of every product classification.
Diets high in Nova 4 foods often also differ in sugar, sodium, fiber, energy density, eating rate, availability, and social context. Statistical adjustments reduce some observed differences but cannot guarantee that all confounding is removed. Avoid attributing the result to one essential quality called “ultra-processing.”
Controlled trials provide short-term causal evidence for specific diets, particularly regarding energy intake and body weight. They do not retroactively validate every Nova 4 boundary or show that each additive or process has the same effect.
8. Opposing 2026 perspectives debate the concept’s use, not proven fraud
A Nature Food perspective published in June 2026 defended Nova’s clarity, validity, and usefulness. Its authors acknowledged that ultra-processed foods are multidimensional—composition, ingredients, processes, and additives—but argued that this complexity does not eliminate the classification’s analytical or predictive value for research and policy. Khandpur et al., 2026 — defense of Nova’s validity and usefulness
A critical perspective in the same issue instead emphasized problems in scientific communication and interpretation of the evidence base. The coexistence of these articles does not constitute a “Nova scandal.” It shows that researchers continue to debate the level at which the category remains useful and how much precision is needed before designing highly targeted policies around it. Critical perspective, 2026 — research and communication about ultra-processed foods
As of July 26, 2026, the sources reviewed here documented no fraud, data fabrication, or major retraction that invalidated Nova. The genuine controversies concern definitions, methods, mechanisms, Group 4 heterogeneity, and translation into recommendations. Calling this a scandal for attention would distort the debate.
9. Conflicts of interest should be described on all sides
Industry funding can create a risk of bias; longstanding involvement in designing or promoting a framework can also create intellectual allegiance. Neither factor alone proves that a result is false. Protocol, data, analyses, limitations, replications, and reporting should be considered together.
The Braesco study disclosed its indirect partial funding and potential food-industry ties. The favorable 2026 perspective disclosed public or nonprofit advisory activities related to nutrition and Nova while declaring no food-industry conflicts. This information helps readers understand the authors’ positions; it does not replace examination of their arguments.
Honest editorial practice cites the study, reports the figures and limitations, and makes the context visible. Accusing one side of being bought or the other of activism resolves neither the classification of an ambiguous yogurt nor the accuracy of a food questionnaire.
10. Use Nova with a simple protocol and an “uncertain” outcome
For each product, retain the exact ingredient list and date. Apply the same criteria to every comparison, document the markers, and use “undetermined” when origin or process information is missing. To track a diet, observe the share and frequency of clearly classified products instead of assigning false precision to a borderline case.
Never decide from the group alone. Add serving size, calories, protein, fiber, sugar, saturated fat, salt, and the product’s role in the meal. A convenient Nova 4 option may sometimes be better than skipping a meal or losing a protein source; a minimally processed option is not automatically suitable in unlimited amounts.
Nalko does not certify ambiguous cases. The app uses available groups, treats simple foods in its database as Nova 1, and aggregates calculable calories from other entries without a group under “Nova undetermined.” Its internal average covers calories associated with Groups 1–4; this tracking remains descriptive.
- Same version of the criteria for all products.
- Complete recipe and date preserved.
- Category recorded as certain, probable, or undetermined.
- Parallel portion and nutritional profile analysis.
- Decision based on several days, not one isolated food.
Frequently asked questions about Nova reliability
Is the Nova classification scientifically validated?
It is widely used and studied, but validity is not binary. Reproducibility often becomes acceptable with detailed data and a structured protocol; its boundaries and validity against an independent reference framework remain variable.
Why do two studies find Nova agreements so different?
Braesco et al. tested a large group of specialists using some ambiguous foods and obtained κ values of 0.32–0.34. Sneed et al. trained and certified coder pairs under a structured protocol and obtained 88.3% agreement and κ 0.75. The populations, data, and methods differed.
What does ANSES criticize about Nova?
ANSES considers the definition non-consensual, some criteria subjective, and the system more focused on formulation than on processes. It also highlights non-exhaustive lists and the absence of a hazard threshold, and recommends examining possible mechanisms separately.
Does good reproducibility prove that Nova predicts health?
No. It shows only that coders apply the rules consistently. Health prediction requires additional studies, while causal interpretation requires separating nutritional profile, amount, food matrix, eating rate, additives, and context.
Do industry ties invalidate the Braesco study?
Not automatically. Funding of less than 5% through an organization that was partly industry-supported and the authors’ declared ties should be visible. The study’s value then depends on its protocol, data, limitations, and replication.
Is there a scientific scandal surrounding Nova?
As of July 26, 2026, the reviewed sources documented no systemic fraud, fabrication, or major retraction invalidating Nova. Genuine scientific controversy remains over criteria, reproducibility, mechanisms, and policy.
Can you use Nova to choose your foods?
Yes, as an additional question about formulation, using consistent rules and an uncertain category. Always consider serving size, nutritional profile, frequency, budget, and the product’s role in your diet as well.
Sources and references
- Monteiro et al., 2016 — formalization of Nova
- Monteiro et al., 2019 — Nova 4 practical criteria
- Braesco et al., 2022 — agreement between specialists
- Braesco et al. — French version with funding and links of interest
- Sneed et al., 2023 — trained coding of food recalls
- SACN, 2023 — institutional review of classifications and their reliability
- ANSES, 2024–2025 — analysis of Nova and the health evidence
- 2026 systematic review — reliability and validity of processing classifications
- Khandpur et al., 2026 — methodological defense of Nova
- Critical perspective, 2026 — scientific communication about ultra-processed foods