Robustness and Explainability of Artificial Intelligence: From technical to policy solutions
Robustness and Explainability of Artificial Intelligence From technical to policy solutions
This technical report by the European Commission Joint Research Centre examines the robustness and explainability of artificial intelligence systems, bridging technical machine learning research with policy and regulatory considerations. It details vulnerabilities such as data poisoning and adversarial examples, discusses transparency and data protection issues under the GDPR, and puts forward policy recommendations for standardization and certification tools for AI.
Summary and content fields are drawn from the document text and are pending editorial review.
Who issued it
Who it governs
Policy makers, regulators, and researchers looking to understand the technical limitations, risks, and governance frameworks surrounding AI robustness and explainability.
What the document contains
The report covers the technical background of artificial intelligence and machine learning, focusing on transparency, model reliability, vulnerabilities like adversarial examples and data poisoning, data protection, and potential policy solutions including certification, impact assessments, and standardization.
The report does not provide binding legal regulations or a final operational standard, but rather offers technical analysis and policy-oriented recommendations for decision-makers.
- AI systems present significant vulnerabilities, such as data poisoning and adversarial examples, that threaten their reliability and security.
- The opaqueness of complex machine learning models creates challenges for transparency and compliance with data protection frameworks like the GDPR.
- Robustness and explainability are closely intertwined, as understanding a model's mechanisms helps assess its failure risks and ensure trustworthy deployment.
- Policy solutions should include impact assessments modeled on Data Protection Impact Assessments (DPIAs), alongside standardized robustness testing and certification schemes.
- Establishing threat-driven procedures and publishing good practices are essential for building trust and managing risks associated with AI deployment.
· Executive summary · Abstract · 1 Introduction · 2 Background: Perspectives from Society and Policy making · 3 Artificial Intelligence and its Robustness and Explainability: Overview and limitations · 4 From technical to policy solutions · 5 Conclusion · References
How it scores
Each score is backed by a quote verified against the document text. Editorial review has not yet confirmed them.
Fairness
2/5Non-discrimination and equitable treatment
diversity, non-discrimination and fairness: accessibility, lawfulness;
2.3 From Data Governance to AI Governance
screening score · 2026-10-23
Human Dignity
2/5Respect for the inherent worth of persons
Users must be guaranteed that their fundamental rights are protected and that products and services are safe to use.
4 1 Introduction
screening score · 2026-10-23
Human Oversight
3/5Humans remain in control of consequential decisions
For critical contexts where no safe and secure technical solution exists, human supervision and final decision -making should remain the default option.
4.1.1 Impact assessments of AI systems
screening score · 2026-10-23
Privacy
3/5Protection of personal data and private life
The General Data Protection Regulation (GDPR), applicable since 2018, has introduced a set of rights that relate to the explainability of AI
1 Executive summary
screening score · 2026-10-23
Transparency
4/5Openness about how AI systems work and are used
promoting transparency in the conception of machine learning models, emphasizing the need of an explainability-by-design approach for AI systems
1 Executive summary
screening score · 2026-10-23
Enforcement
2/5Consequences and compliance mechanisms with real force
This characteristic raises the question of a possible expiration date for the certification.
4.1.2 Testing
screening score · 2026-10-23
Impact Assessment
4/5Required assessment of an AI system's risks/impacts before or during use
developing a methodology to evaluate the impacts of AI systems on society built on the model of the Data Protection Impact Assessments (DPIA) introduced in the GDPR
1 Executive summary
screening score · 2026-10-23