AI reliability
The ability of an AI system to perform as required, without failure, over a specified time interval and under defined conditions. Reliable AI systems consistently operate correctly, promoting user trust and reducing risk in critical applications.
AI resilience
The capacity of an AI system to maintain functionality and performance in the face of unexpected challenges, including adversarial attacks or environmental shifts. Resilient AI systems are robust and recover gracefully from disruptions.
AI risk management
Coordinated activities to identify, assess, and mitigate risks throughout the AI lifecycle. AI risk management enhances understanding of potential impacts, reduces harm, and improves the trustworthiness of AI systems across contexts.
AI robustness
The ability of an AI system to sustain performance across a range of scenarios, including those not initially anticipated. Robust systems demonstrate reliability and minimize harm, even in varied or unexpected environments.
AI safety
Assurance that an AI system will not, under defined conditions, cause harm to human life, health, property, or the environment. Safe AI systems incorporate design practices, rigorous testing, and controls to avoid dangerous states or failures.
Adversarial Attacks
Techniques used to exploit vulnerabilities in AI systems by manipulating data or inputs, causing the AI to perform unintended actions or produce incorrect outputs. Resilient AI systems are designed to withstand such attacks and maintain functionality.
Algorithmic Auditing
A thorough examination of algorithms within AI systems to assess their accuracy, fairness, and compliance with legal and ethical standards. Algorithmic audits aim to identify and mitigate potential biases and ensure responsible AI deployment.
Algorithmic Bias
A tendency of an AI algorithm to produce discriminatory outcomes, often due to skewed training data or flawed model design. Algorithmic bias audits address disparities, ensuring outcomes align with fairness principles and anti-discrimination laws.
Algorithmic Fairness
A principle ensuring AI systems do not perpetuate or amplify existing biases, particularly against protected groups. This includes detecting and mitigating discriminatory outcomes to align with fundamental rights and anti-discrimination laws.
Algorithmic Transparency
The principle that AI systems’ decision-making processes should be clear and understandable, enabling users to comprehend how outputs are derived. Transparency fosters accountability and builds user trust by providing insight into AI operations.
Automated AI Compliance
The use of AI-driven tools to automatically monitor and enforce compliance with regulatory, ethical, and operational standards. Automated compliance tools ensure continuous adherence to requirements without manual oversight.
Automated Decision-Making
Processes in which AI systems independently make decisions or recommendations impacting individuals or groups. Automated decisions must align with fairness, transparency, and accuracy standards, particularly in high-risk applications.
Biometric Categorization
The process of assigning individuals to specific groups or categories based on biometric data, such as age, gender, or behavior, excluding any direct identification purposes or consent-based verification processes.
Biometric Data
Data derived from physical, physiological, or behavioral characteristics enabling unique identification of a natural person, including facial images, fingerprints, and gait, processed for either identity verification or categorization purposes.
Biometric Identification
The automated process of identifying individuals by comparing their biometric data with reference data, typically without active participation from the individual, to establish identity in various environments, including security settings.
Black Box AI Model
AI models with complex internal operations that are opaque or difficult to interpret by users. These models pose challenges for transparency and accountability, often requiring explainability measures to ensure trust and understanding.
Business Continuity Plan
A strategic approach to maintaining AI system functionality during disruptions, ensuring ongoing operations. Plans include measures for data backup, disaster recovery, and contingency actions to mitigate risks of operational interruptions.
Business Impact Analysis
An assessment process evaluating the potential effects of AI system failure or disruption on business functions. This analysis identifies critical operations, quantifies impacts, and guides risk management and continuity planning efforts.
Compliance Documentation
Records and materials that detail an AI system’s design, functionality, and risk assessments, ensuring conformity with applicable EU regulations, including technical specifications, testing protocols, and operational guidance for responsible use.
Compliance Mechanism
Structured approaches within an organization to ensure AI systems align with applicable legal and regulatory requirements. Compliance mechanisms encompass policies, processes, and monitoring activities to manage AI risks effectively.
Conformity Assessment
A systematic evaluation process to verify that high-risk AI systems meet all EU compliance requirements, ensuring their safe deployment and adherence to standards that protect users and affected persons’ health, safety, and fundamental rights.
Continuous Monitoring
Ongoing observation and assessment of an AI system’s performance, identifying and addressing emergent risks and system failures over time. Continuous monitoring supports adaptation to evolving conditions and stakeholder needs.
Cybersecurity
Measures implemented to protect AI systems from unauthorized access, manipulation, or malicious attacks. These measures ensure the integrity, confidentiality, and reliability of the AI system’s performance across its operational lifecycle.
Data Governance
Policies and procedures ensuring data used in AI systems is accurate, representative, and privacy-compliant, particularly in high-risk contexts. Effective data governance prevents biases, protects user data, and maintains the integrity of AI processes.
