Artificial intelligence and nutrition in the EU: three frameworks, one product

The regulatory landscape

The EU AI Act and the risk-based approach

Regulation (EU) 2024/1689, known as the EU AI Act, entered into force on 1 August 2024 and becomes fully applicable on 2 August 2026. It introduces a horizontal legal framework for artificial intelligence in the European Union based on levels of risk: prohibited, high-risk, limited-risk, and minimal-risk.

For nutrition applications, the decisive question is intended purpose. A consumer app that provides general lifestyle guidance will often fall into the limited-risk category. That same system may move into high-risk territory if it functions as part of a regulated medical product, or if it supports decisions connected to diagnosis, prevention, or treatment. The boundary matters because high-risk classification brings a much heavier compliance burden, including technical documentation, quality systems, data governance, conformity assessment, and post-market monitoring.

Even where a system does not qualify as high-risk, the AI Act still matters. User-facing systems must meet transparency requirements, and design choices around oversight, accountability, and user protection remain relevant

Nutrition and health claims regulation

Regulation (EC) No 1924/2006 sets the legal framework for nutrition and health claims made on foods and food supplements in the EU. The core logic is well established: only permitted nutrition claims and authorised health claims may be used, and only under the relevant conditions of use. Claims must be clear, accurate, and supported by accepted scientific evidence.

For traditional food communication, the framework is familiar. For AI-generated outputs, it becomes more complex. A conversational interface can generate statements that function as health claims even if they were never pre-approved by a regulatory or marketing team. If a chatbot tells a user that a nutrient "supports immune function" or suggests that a botanical ingredient "reduces cardiovascular risk", the legal analysis does not disappear simply because the statement appeared in a personalised dialogue rather than on-pack.

This is why the EU Register of nutrition and health claims remains a critical operational reference, not only for labels and advertising, but increasingly for digital products and dynamic content systems.

EFSA’s scientific expectations

EFSA remains central to the European evidence ecosystem in nutrition and food regulation. Its scientific opinions, methodological standards, and approach to substantiation continue to shape the benchmark for what counts as sufficiently supported in the context of health-related communication.

In parallel, EFSA has been developing its own structured approach to artificial intelligence. Its AI roadmap and related initiatives point to the use of AI in literature discovery, evidence organisation, and systematic review support. The underlying message is consistent: AI can improve the way evidence is identified and processed, but it does not replace scientific judgement. Final assessment remains expert-led.

That principle is highly relevant for private-sector nutrition AI. Any company presenting its outputs as evidence-based, scientifically grounded, or aligned with EFSA logic should be able to show where its evidence comes from, how it is curated, how its models are validated, and where human review remains in the loop.

The main use cases

Personalised recommendation engines

This is the most visible category in consumer nutrition AI. These systems collect user inputs such as age, goals, dietary preferences, habits, and sometimes data from trackers or glucose sensors, then generate dietary advice, meal suggestions, nutrient targets, behavioural nudges, or supplement recommendations.

From a regulatory perspective, these tools sit in a sensitive space. Under the AI Act, they may remain limited-risk if they are clearly framed as wellness tools. Under the claims framework, however, they can still create exposure if their outputs imply specific health benefits that go beyond authorised wording or established conditions of use.

Scientific substantiation also matters here. If the recommendation logic is presented as personalised and evidence-based, it should rest on credible references such as dietary reference values, EFSA scientific opinions, and recognised nutrition science, rather than opaque or selectively chosen literature.

Nutrition based on biomarkers or genetic data

A more advanced class of services integrates laboratory data, microbiome profiles, or genetic tests with algorithmic interpretation to generate tailored nutrition advice. These products are often presented as highly innovative, but they also move much closer to the medical boundary.

Once an AI system interprets biological markers in a way that shapes recommendations linked to disease prevention, management, or treatment, the regulatory position changes significantly. At that point, the system may fall within the scope of MDR or IVDR and therefore within the high-risk regime of the AI Act.

Communication becomes equally delicate. There must be a clear separation between general health maintenance and disease-related messaging. A product cannot quietly borrow the language of clinical utility while still expecting to remain in the lighter wellness category. If a company chooses to operate near that boundary, it needs a much stronger compliance strategy and a much stronger evidence package.

B2B evidence engines for product development

A third category includes internal or business-facing tools that structure scientific literature, EFSA opinions, regulatory databases, and claims information to support formulation, substantiation strategy, or regulatory decision-making. These systems may not produce consumer-facing text, but they can still have substantial downstream impact.

This is especially relevant for platforms like Nutri-AI. A well-designed evidence engine can help reduce noise, improve traceability, and anchor decision-making in authoritative sources. But the quality of the output depends entirely on the quality of the evidence pipeline. If outdated studies, non-authoritative databases, or poorly filtered claims information are fed into the system, the platform may simply scale bad regulatory judgement more efficiently.

From an AI Act perspective, many of these tools may remain outside the high-risk category if used internally. Still, their strategic importance is high, especially when they become the basis for product positioning, claim selection, or risk assessment.

The grey zones that matter most

Wellness versus medical decision support

This is the most important grey zone in the field. A chatbot that helps users eat more vegetables or improve meal balance is one thing. A system that adjusts supplement dosages for individuals with specific conditions based on blood markers is something else entirely.

The difficulty is that many products are built in the space between the two. Their functional design may suggest medical relevance, while their branding remains carefully anchored in the language of lifestyle and wellness. That mismatch creates risk. Regulators will not look only at the interface. They will look at intended purpose, functionality, claims, and the overall impression created by the product.

For founders and product teams, the lesson is straightforward: define intended purpose precisely, document it clearly, and ensure that product design, user messaging, and scientific logic all align with that positioning.

Dynamic claims generated in real time

Generative systems introduce a challenge that classic regulatory frameworks did not fully anticipate: claims can now emerge dynamically, in real time, in response to user prompts. A model may produce non-compliant wording even when the product team never explicitly scripted it.

That makes governance a technical issue as much as a legal one. If a model is allowed to speak freely about ingredient benefits, it may drift into unauthorised or disease-related wording. In the nutrition space, that is not a minor drafting problem. It is a compliance risk built into the architecture of the system.

The most credible mitigation strategies are therefore not purely editorial. They are structural: constrained generation, authorised claim libraries, rule-based validation layers, and logging systems that connect outputs back to approved evidence and wording logic.

Scientific substantiation in an AI-enabled environment

AI can accelerate evidence review dramatically. It can screen literature at scale, organise documents, detect patterns, and help identify relevant sources far more quickly than manual workflows alone. But speed is not the same as scientific quality.

Evidence pipelines built on AI still need methodological discipline. They require source validation, relevance criteria, documentation of inclusion logic, review against trusted standards, and expert interpretation. Otherwise, the system may produce outputs that look sophisticated while resting on weak or non-transparent foundations.

This is where EFSA’s approach remains instructive. AI can support the evidence process, but scientific accountability remains human. For any company operating in nutrition AI, that principle should not be treated as a philosophical preference. It should be treated as an operational rule.

What this means in practice

A three-axis mapping exercise from the start

For every product feature, companies should assess three questions in parallel.

• Where does the feature sit under the AI Act, based on intended purpose and any link to MDR, IVDR, or Annex III use cases?

• Could the output be interpreted as a nutrition or health claim under Regulation (EC) No 1924/2006?

• What is the evidence base behind the output, and does it meet a defensible scientific standard?

This kind of mapping is most useful when it happens early. It is difficult to retrofit regulatory clarity onto a product whose positioning, logic, and content model were never designed with these questions in mind.

Compliance by design

For nutrition AI, compliance should be treated as an architectural choice, not a late-stage control function.

In practical terms, that means:

• constraining generative systems to approved claims logic where relevant

• maintaining clear separation between educational content, product claims, and medical content

• documenting evidence sources and update logic

• ensuring traceability between recommendations and their underlying data or rule sets

• building expert review into significant changes to the recommendation engine or evidence base

These are not simply defensive measures. They are part of what makes a product trustworthy.

Due diligence and investment readiness

For investors, nutrition AI can no longer be evaluated on product appeal and growth potential alone. Regulatory maturity is becoming part of the commercial quality of the asset.

A serious diligence process should test whether the company has defined intended purpose clearly, analysed the wellness versus medical boundary, built a coherent claims strategy, and anchored its evidence workflows in authoritative sources such as EFSA opinions and official reference values.

Companies that can demonstrate this level of regulatory and scientific discipline are likely to be better positioned as scrutiny increases and enforcement becomes more systematic across digital health, food, and AI markets.

#NutriAI #NutriAINewsletter #ArtificialIntelligence #AI #Nutrition #ScientificCommunication #FoodTech #FoodSafety #AIRegulation #EFSA #RegulatoryCompliance #ISO42001 #HealthClaims #DigitalInnovation #ResponsibleAI #AITransparency #Governance #DataScience #FoodCompliance #DigitalNutrition #FoodLaw #HighRiskAI #TrustInAI #AINews #ScientificCommunication #EUAIAct #MedicalEducation #AIliteracy #ContinuingEducation #Dietitians #Nutritionists #LargeLanguageModels #AIAct #ClinicalDecisionSupport #DigitalHealth #Nutrition

Disclaimer: All rights to images and content used belong to their respective owners. This article is provided for educational and informational purposes only. It does not constitute legal or regulatory advice. Organizations should consult qualified legal and regulatory experts before implementing AI systems in the nutrition sector.

--------------------------------------------------------------------------

Bibliographic and Regulatory References

• Regulation (EU) 2024/1689 (EU AI Act). In force from 1 August 2024; fully applicable from 2 August 2026. European Commission

• Regulation (EC) No 1924/2006 (NHCR). Nutrition and health claims made on foods. Consolidated text

• EU AI Act, Annex III. High-risk AI systems under Article 6. Official text

• EFSA (2022). Update on EFSA AI roadmaps about risk assessment (EN-7339). EFSA PDF

• EFSA (2025). Programming document 2025–2027. EFSA PDF

• EFSA. AI@EFSA, dedicated page on EFSA’s exploration of AI since 2017. efsa.europa.eu

• Barizzone F. et al. (2024). Introducing AI in EFSA systematic reviews. ECETOC workshop. ECETOC PDF

• De Groote W. et al. (2024). The EU Artificial Intelligence Act (2024): implications for healthcare. Health Policy, 149. ScienceDirect

• Klonner M. et al. (2024). Navigating the European Union Artificial Intelligence Act for healthcare. npj Digital Medicine, 7:218. Nature

• Osborne Clarke (2025). European Commission consultation shapes high-risk AI classification in life sciences. osborneclarke.com

• DPO Consulting (2025). High-risk AI systems under the EU AI Act. dpo-consulting.com

• UK Government / FSA (2021). Nutrition and health claims: guidance to compliance with Regulation (EC) 1924/2006. gov.uk

• University of Parma / foodforfuture (2024). Botanical substances and health claims: an awaited judgment. foodforfuture.unipr.it