The Revolution of Human-AI Partnership in European Food Security

The digital transformation sweeping through the European Food Safety Authority (EFSA) represents a paradigm shift in the approach to food risk assessment. Through the strategic implementation of human-centric artificial intelligence technologies, EFSA is redefining global standards for regulatory science, establishing a new model of collaboration between human expertise and advanced computational capabilities that promises to transform not only the European landscape but also the entire global food safety ecosystem.

Alessandro Drago

11/3/2025

Automation vs Augmentation: Two Paradigms in Comparison

Before analyzing EFSA's revolutionary approach, it is fundamental to understand the theoretical and practical distinction between automation and augmentation in the context of artificial intelligence applied to scientific and regulatory processes. This distinction is not merely academic, but represents a strategic choice that determines AI's impact on the organization, workers, and society as a whole.

Fundamental Definitions and Characteristics

Automation implies that machines completely assume control of a human task, replacing human intervention with algorithmic processes. In this paradigm, the objective is to eliminate the need for human supervision, creating autonomous systems capable of executing tasks without external intervention. Automation typically focuses on well-structured activities, with clearly defined inputs and predictable outputs.

Augmentation, on the contrary, means that humans collaborate closely with machines to perform a task. In this model, AI does not replace human judgment but enhances it, expanding cognitive capabilities and allowing experts to process volumes of information previously inaccessible. Augmentation maintains humans at the center of the decision-making process, using AI as a tool for supporting and amplifying existing competencies.

The empirical research by Raisch and Krakowski (2021) identified the "automation-augmentation paradox," demonstrating that in the managerial domain these two AI applications are interdependent across time and space, creating a paradoxical tension that requires sophisticated management approaches. Overemphasizing automation or augmentation feeds reinforcement cycles with negative organizational and social outcomes, while adopting a broader perspective that encompasses both approaches can generate beneficial complementarities for business and society.

Applications in the Scientific-Regulatory Context

In the specific context of regulatory science, the distinction assumes critical connotations. Automation in systematic reviews can include completely automated identification of relevant studies, automatic data extraction without human supervision, and automatic generation of syntheses. While this approach can offer significant reductions in time and costs, it presents substantial risks in terms of accuracy and reliability, particularly problematic in high-risk contexts such as food safety.

Augmentation in systematic reviews, instead, uses AI to assist human researchers in specific tasks while maintaining expert supervision. This includes assisted screening where AI prioritizes documents but experts maintain final decision-making control, semi-automated data extraction with human validation, and cognitive analysis that identifies patterns to support expert interpretations.

A recent study on AI applications in rheumatology demonstrated that the augmentation approach can reduce screening times by 83% while maintaining identification of 95% of relevant articles, highlighting how human-AI collaboration surpasses both completely manual and completely automated approaches.

Implications for Trust and Responsibility

The choice between automation and augmentation has profound implications for public trust and scientific accountability. As highlighted by Vaccaro et al. (2024), systems that combine human intelligence and AI tools can address questions of social importance but present specific challenges. The research identified that human augmentation exists when human-AI groups perform better than humans working alone, while human-AI synergy requires that they perform better than both humans and AI operating individually.

In the regulatory context, augmentation offers critical advantages in terms of transparency and explainability. When human experts maintain final decision-making control, they can provide justifications for decisions made, identify limitations of AI systems, and intervene when necessary. This is particularly important in high-risk domains where decisions have direct consequences on public health.

Empirical Evidence and Comparative Performance

The scientific literature provides growing evidence of the superiority of the augmentation approach in knowledge-intensive contexts. Marguerit (2025) analyzed AI effects on the US job market from 2015 to 2022, distinguishing between automation AI and augmentation AI. The results support the hypothesis that automation AI technologies are more applicable to low-skilled occupations, while augmentation AI is more effective for high-skilled roles that require creativity, critical judgment, and interpersonal skills.

In healthcare, a comparative analysis between data scientists and clinicians revealed that both groups prefer levels 2 and 3 of human-AI teaming (where AI augments human performance or humans augment AI performance) for 4 out of 6 core tasks in intensive care units. Only for monitoring was level 4 preferred (AI performing without human input), highlighting how even in highly technological contexts collaboration remains preferable to complete automation.

The Strategic Framework: From Vision to Implementation

The AI@EFSA initiative, formalized in the 2022 roadmap, represents not simply a technological upgrade, but a structural transformation that aims to achieve three fundamental objectives by 2027: increase accessibility of the scientific evidence corpus, expand its breadth, and enhance the reliability of the risk assessment process. This vision is supported by a robust architectural framework that rests on four critical technological pillars: reliable AI governance, sophisticated ontologies for data organization, big data infrastructure, and comprehensive Machine Learning Operations (MLOps) capabilities.

The implementation model adopted by EFSA is distinguished by its human-centric approach, where artificial intelligence does not replace human scientific judgment, but enhances it through augmentation processes rather than complete automation. This paradigm, defined as a "human-augmented approach," maintains final decision-making responsibility with scientific experts, while ensuring significantly increased operational efficiency.

The AI task force established in January 2024 operated on three strategic dimensions: development of an internal network of interested experts, identification of IT investments necessary for AI adoption, and support for defining internal guidelines compliant with recommendations from EU control and audit bodies. By the end of 2024, the task force reached consensus on a set of guiding principles and preliminary definitions for EFSA's AI approach, providing a compass to guide future decision-making processes.

Quantitative Performance: Results of the Transformation

Empirical data collected from the pilot implementation of AI technologies at EFSA document extraordinary results in terms of effectiveness and acceptance by the scientific community. Analysis of 41 participants using AI-assisted systematic reviews revealed exceptional satisfaction rates: 21 completely satisfied users and 14 partially satisfied users, corresponding to a 95% satisfaction rate.

Efficiency gains manifest through precise quantitative metrics: 90% of users experienced significant time savings, while 90% registered optimizations in human resource allocation. Independent validation conducted by the UK government's Behavioural Insights Team confirmed that AI-assisted systematic reviews are completed with a 23% reduction in time compared to traditional approaches.

Granular analysis of the systematic review process reveals where AI generates maximum impact: 30% time reduction in screening phases, 56% improvement in data analysis, and 43% enhancement in synthesis phases. These metrics represent transformative gains that free researchers for high-value-added scientific activities, enabling them to focus on critical thinking, contextualized interpretation, and elaboration of complex judgments that require true understanding of scientific and social implications.

EFSA's performance in 2024 recorded scientific production of 550 questions, with significant improvements in the timeliness adoption index compared to 2023, despite the greater scientific and procedural complexity of the Authority's work. This result testifies to the effectiveness of the AI-assisted approach in managing growing volumes of scientific work while maintaining high quality standards.

DAISY and DistillerSR: Technological Excellence in Action

The heart of EFSA's AI implementation is represented by the DistillerSR system, enhanced by the DAISY platform (DistillerSR Artificial Intelligence System). This integrated technological solution systematically manages the most labor-intensive phases of systematic reviews, including keyword identification, deduplication, reference prioritization, title and abstract screening, automatic classification, and screening error verification.

DAISY system performance is documented through rigorous metrics: AI achieves an average sensitivity of 90% in abstract screening, with precision, recall, and F1-scores oscillating between 80% and 100% in data extraction tasks. In the specialized context of risk assessment, where omission of a single relevant study can influence critical decisions for public safety, these performance metrics represent the difference between good science and scientific excellence.

The pilot implementation of DAISY through five specific use cases demonstrated the effectiveness of the modular approach: from evaluating the usability of automatic summarization tools for public consultations, to automating keyword identification in systematic review processes, to creating personalized manuals to facilitate AI-assisted relevance screening. This applicative diversification highlights the versatility and adaptability of the developed technological framework.

Risk Governance: Validation and Scientific Responsibility

EFSA's approach to AI implementation is distinguished by sophisticated understanding of current technological capabilities and limitations. The implementation strategy focuses on high-volume, low-subjective-judgment tasks where AI excels: terminological normalization, duplicate removal, relevance screening, and literature clustering.

The implemented governance framework ensures that every AI application is subjected to rigorous validation against gold standards and continuous expert supervision. This control architecture ensures that AI serves as an enhancement tool rather than replacement, maintaining final scientific responsibility always with human experts. Standard Operating Procedure (SOP) 048 "Govern Technology and Transformation" provides the methodological framework for managing IT processes according to market best practices, establishing policies and procedures that ensure strategic alignment between business and technology.

The Web Application for Distiller Model Control, developed specifically for EFSA, represents an additional level of quality assurance, providing modules for generating random samples of references for AI tool training and for evaluating performance on validation sets. This continuous control system ensures that scientific quality remains the guiding parameter of all AI implementations.

Ontologies and Knowledge Networks: The Architecture of Knowledge

EFSA's AI implementation rests on a sophisticated ontological architecture that represents one of the fundamental pillars of digital transformation. The survey conducted on 40 ontologies and 7 knowledge organization systems related to food safety provided the basis for developing an integrated semantic framework. This systematic work identified the most relevant ontologies for the food safety domain, categorizing them according to criteria of relevance, coverage, semantic coherence, and reusability.

The FSO (Food Safety Ontology) project developed within the DiTECT program illustrates the methodological approach adopted: maximum reuse of existing and widely used ontologies, such as AGROVOC for the agricultural domain, CHEBI for chemical compounds, and SOSA for sensors and observations. This semantic integration strategy enables interoperability between different scientific domains, facilitating cross-analysis of heterogeneous data.

The ontological architecture supports advanced functionalities such as knowledge mining and "Chat with your documents" scenarios that are revolutionizing how scientists interact with research literature. These cognitive applications allow discovering patterns and relationships in millions of texts that would require decades of manual work to be identified and digested by human researchers.

Cognitive Analytics and Machine Learning: Towards Augmented Intelligence

EFSA's strategic roadmap incorporates advanced cognitive analytics, including machine learning and natural language processing, to discover patterns and relationships in information from millions of texts, books, online articles, and other sources, including social media. This cognitive analysis capability represents a qualitative leap in scientific evidence processing, allowing processing of information volumes that exceed traditional human capabilities by orders of magnitude.

The AI4NAMs (Artificial Intelligence for New Approach Methodologies) project systematically explored the potential of applying AI methods to NAMs data collection and integration, supporting better chemical risk assessment while simultaneously reducing animal testing. Results demonstrated that AI can effectively support the entire NAMs data management workflow, from research and extraction to harmonization and integration into networks similar to Adverse Outcome Pathways (AOP).

Fine-tuning Large Language Models for food and feed safety represents a particularly promising frontier. Adapting pre-trained linguistic models to food safety domain specificities enables developing specialized conversational assistants capable of supporting experts in analyzing and interpreting complex technical documents.

International Collaborations and Strategic Partnerships

EFSA's collaborative model extends through a sophisticated network of international and inter-agency partnerships. The Cross-agency One Health Task Force, established by five EU agencies (ECDC, ECHA, EEA, EFSA, and EMA), released the Framework for Action to operationalize the One Health approach through coordination, research promotion, capacity strengthening, communication, and joint activities.

The collaboration framework provides for introducing AI into European agencies through collaboration and mutual sharing with EU and international bodies. This inter-agency vision aims to establish a common European AI roadmap, identifying and implementing shared interest use cases and creating a common governance structure for AI applications among European agencies.

The FoodSafety4EU project represents a concrete example of this multi-stakeholder collaboration, developing a platform that connects food safety system actors to facilitate access, sharing, and exchange of scientific knowledge, resources, and data. The platform supports a participatory multi-actor approach that experiments and enables a proactive and co-creation approach for food safety system transformation.

Agile Methodology and Digital Transformation

Adopting Agile methodology represents a paradigm shift in EFSA's approach to digital transformation. Introducing agile methodologies for faster improvements and greater customer focus is supported by a governance framework that balances operational flexibility and strategic control. Projects implementing the 2027 work program address change management and development of digital skills necessary to exploit 21st-century tools.

The Semi-Agile methodology described in EFSA's standard operating procedures provides a structured framework for managing technological transformation projects. This approach allows rapid response to requirement changes while maintaining traceability and quality control necessary in a regulatory environment.

Enhancing business analytics capabilities and developing interactive dashboards support data-driven decision-making at all organizational levels. These tools enable real-time monitoring of AI-assisted process performance and rapid identification of improvement areas or potential criticalities.

Cybersecurity and Data Management

EFSA's digitalization strategy recognizes cybersecurity as a critical element for AI implementation success. Strengthening cybersecurity across all EFSA operations is identified as a strategic priority in the Strategy 2027 mid-term review. The increasing interconnectivity of AI systems requires sophisticated security protocols to protect data integrity and prevent attacks that could compromise scientific assessment reliability.

Implementing FAIR (Findable, Accessible, Interoperable, Reusable) frameworks for data management ensures that datasets used for AI model training and validation are high quality and free from systematic biases. The data flow mapping methodology developed by EFSA provides comprehensive understanding of information pathways at the national level, identifying existing challenges and implementable common solutions.

Data Protection Impact Assessment (DPIA) is systematically applied before introducing new technologies critical for personal data processing, ensuring GDPR compliance and maintaining public trust in responsible AI use.

Case Studies and Sectoral Applications

AI implementation at EFSA manifests through specific case studies that demonstrate practical applicability of developed technologies. AI-driven monitoring of front-of-package information on food products represents an example of how artificial intelligence can automate traditionally labor-intensive processes. Applying computer vision and natural language processing techniques to recognition of certification marks and nutritional information on infant foods demonstrated the effectiveness of the automated approach in market data collection.

Predicting compliance for food safety through machine learning approaches illustrates how AI can support more targeted and effective inspections. Using synthetic minority oversampling technique together with random forest algorithms achieved 84% detection performance for non-compliant outlets, demonstrating AI's potential in improving official control efficiency.

The agricultural productivity transformation project through AI-driven predictions highlights artificial intelligence applicability along the entire food chain. Integrating AI tools in food processing, including plasma, ozone, and Pulsed Electric Field technologies, shows how technological innovation can simultaneously improve safety and production efficiency.

Future Perspectives: Vision 2027 and Beyond

EFSA's strategic roadmap extends toward advanced technological scenarios that will revolutionize interaction between scientists and research literature. Planned developments include "Chat with your documents" functionalities that will enable conversational interrogations of the scientific corpus, cognitive analytics for pattern discovery in millions of texts, and systematic review automation that would require decades of manual work.

The international collaboration framework provides for extending partnerships to agencies such as ECHA, EMA, and other EU bodies, with creation of "Academy" model networks for capacity building and knowledge exchange. This collaborative vision could transform regulatory science not only in Europe, but establish global standards for responsible AI integration in scientific research.

Integration with the EU common data platform on chemicals represents a strategic challenge and opportunity for EFSA. Collaboration with DG Health and Food Safety, other EU agencies, and the Joint Research Centre in structuring and automating submission, validation, and evaluation of regulatory dossiers will require significant investments in AI and data science expertise.

Challenges and Risk Mitigation

AI implementation at EFSA is not without significant challenges. The need to balance technological innovation with ethical considerations requires a prudent and methodical approach. Issues related to data ownership, confidentiality requests from applicants, and the need for transparency toward Member States represent tensions that must be managed through sophisticated governance frameworks.

Data quality and availability represent fundamental prerequisites for AI application success. Many food safety authorities, especially in middle and low-income countries, face data scarcity and capacity constraints that limit adoption of advanced AI technologies. Strengthening AI and data management literacy, particularly in the public sector, will be key to unlocking these technologies' benefits.

The vulnerability of food businesses to AI products and applications that are not rigorously tested or validated represents an emerging systemic risk. The UK FSA has identified the need for independent validation mechanisms and standards to ensure that AI systems used in food safety are safe, reliable, and fit for purpose.

Training and Capacity Building

Developing AI competencies represents a critical strategic investment for EFSA transformation success. Continuous training and upskilling of current staff are essential, especially in coming years. By the end of 2027, EFSA aims to have established fundamental capabilities necessary to responsibly integrate AI into the organization, including not only technical tools but also governance and control frameworks.

The open innovation approach suggested for EFSA involves strategic knowledge exchange, integrating external insights and sharing central ideas to accelerate internal innovation. Collaborations with academic, industrial stakeholders and other competent authorities that have already developed strategies or services useful for EFSA's internal processes represent accelerated learning opportunities.

Creating repositories and documentation of applied AI tools, together with comparison and evaluation of techniques used, provides a shared knowledge base that supports standardization of best practices.

Global Impact and Model Sustainability

The EFSA model is already influencing the international approach to food safety. FAO has recognized AI's transformative potential in food safety management, with the document "Artificial intelligence (AI) for food safety – A literature synthesis, real-world applications and regulatory frameworks" documenting adoption of similar approaches by global national authorities.

The FAO publication highlights real AI applications through laboratory testing, inspections and surveillance, border control prioritization, regulatory efficiency, and risk communication, providing the first global overview of AI implementation in the food safety sector. Case studies include emerging examples from middle and low-income countries, demonstrating scalability and adaptability of the conceptual framework developed by EFSA.

The EFSA model's sustainability is guaranteed by its scalable architecture and growing collaboration with academic institutions and international agencies. The FoodSafety4EU and HOLiFOOD projects are already piloting co-developed approaches for AI-driven identification of emerging risks, demonstrating the replicability and adaptability of the EFSA framework. Investment in AI literacy is identified as an essential element to enable informed choices on responsible use of AI tools in food safety.

Ethical Considerations and Social Responsibility

EFSA's human-centric approach to AI implementation reflects deep awareness of ethical and social implications of emerging technologies. The strategic partnership between human and artificial intelligence does not aim to replace scientific expertise, but to amplify it through tools that enable processing volumes of information previously inaccessible.

Final scientific responsibility always remains with human experts, ensuring that critical judgment, contextual interpretation, and understanding of social implications remain at the center of the decision-making process. This fundamental principle distinguishes EFSA's approach from purely automated AI implementations, maintaining the human dimension essential for public trust in regulatory science.

Transparency and accountability represent pillars of AI implementation at EFSA. Every application is documented, validated, and subjected to continuous supervision, creating a system of checks and balances that protects scientific integrity while exploiting advanced technological potentials.

The ongoing transformation in Brussels represents more than a technological evolution: it is the demonstration that when artificial intelligence is implemented with scientific rigor, appropriate governance, and continuous validation, it can dramatically improve scientific efficiency without compromising the integrity that public trust requires. The EFSA model establishes a new paradigm for 21st-century regulatory science, where human expertise and advanced computational capabilities operate in synergy to protect public health with unprecedented precision and efficiency.

This visionary approach not only transforms how EFSA conducts its scientific assessments, but creates a replicable template for regulatory authorities worldwide, demonstrating that technological innovation and scientific excellence can coexist and mutually enhance each other in the supreme interest of global food safety.

#Nutriai #education #learning #science #nutrition #research #health #wellness #foodscience #scientificliteracy #educationalcontent #evidencebased #nutritionscience #publichealth #scicomm #sciencecommunication #educationalpurpose #knowledgesharing #professionaltraining #healthyeating #foodsafety

Disclaimer: This content is used solely for educational and informational purposes. All texts and images are not owned or claimed by the user. No copyright or proprietary rights are claimed or transferred.

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

1. https://journals.aom.org/doi/10.5465/amr.2018.0072

2. https://infomineo.com/artificial-intelligence/automation-vs-augmentation-will-ai-replace-or-empower-professionals-2/

3. https://access.archive-ouverte.unige.ch/access/metadata/dd0713db-1880-4c09-9701-52ea58633532/download

4. https://www.aalpha.net/blog/human-ai-collaboration-augmenting-capabilities-with-agentic-platforms/

5. https://ratio.se/en/publications/artificial-intelligence-and-management-the-automation-augmentation-paradox

6. https://link.springer.com/10.1007/s00607-023-01181-x

7. http://thesai.org/Downloads/Volume15No1/Paper_122-Transformative_Automation_AI_in_Scientific_Literature_Reviews.pdf

8. https://pmc.ncbi.nlm.nih.gov/articles/PMC11826052/

9. https://pmc.ncbi.nlm.nih.gov/articles/PMC6582554/

10. https://pmc.ncbi.nlm.nih.gov/articles/PMC11624811/

11. https://www.nature.com/articles/s41562-024-02024-1