Interaction Transparency

Interaction transparency refers to the clarity and understandability of communication between users and AI systems. It ensures that users are aware of how the AI operates, what inputs it requires, and what outcomes they can expect. This transparency helps users trust AI by making interactions predictable and easily interpretable, often achieved through well-designed user interfaces or clear system explanations.

Interpretability

The ability of AI systems to generate outcomes that can be understood by end users and regulators, especially in high-risk contexts, to promote transparency, accountability, and informed decision-making.

Legal Compliance

The adherence of AI systems to legal standards and regulations, ensuring their deployment aligns with applicable laws governing data protection, safety, and ethical guidelines to protect users and affected individuals.

Limited risk AI systems

AI systems with moderate potential for adverse impact, subject to transparency and information requirements but not classified as high-risk. Users must be informed they are interacting with AI, ensuring responsible and transparent usage.

Machine Learning

An AI technique where systems improve their performance by learning from data without explicit programming. Machine learning models require proper documentation and compliance to minimize risks of bias and errors.

Machine Learning Fairness (ML Fairness)

Techniques and practices aimed at promoting equitable outcomes in machine learning systems by addressing potential biases. ML Fairness seeks to prevent discrimination and ensure AI models are fair across demographics.

Minimal risk AI systems

AI systems with low potential to impact users’ health, safety, or fundamental rights. These systems typically require no specific regulatory oversight and are presumed safe for general use without additional compliance obligations.

Model Accountability

Ensuring AI models operate in line with ethical and regulatory standards, with responsibility for outcomes assigned to providers. Model accountability includes transparency, risk management, and compliance with legal obligations.

Model Generalizability

The extent to which an AI model can maintain performance on new or diverse data that differs from its training data. Generalizable models perform effectively across varied contexts, enhancing reliability and reducing risk.

Model Interpretability

The ability of AI systems to produce results that are understandable to users, enabling them to comprehend the reasoning behind AI-driven outcomes. Interpretability is key for transparency, trust, and effective decision-making.

NIST Compliance Standards

Guidelines established by the National Institute of Standards and Technology for ethical, transparent, and secure AI systems. NIST standards support regulatory compliance, risk management, and protection of user rights.

Post-Market Monitoring

Continuous monitoring of an AI system’s performance and compliance after its deployment, enabling the provider to detect and correct any issues or unintended consequences impacting safety, accuracy, or compliance.

Privacy Impact Assessment

An evaluation to determine how AI systems handle personal data, assessing compliance with data protection standards, including GDPR. The assessment helps identify privacy risks and implement safeguards to protect user information.

Privacy-Enhanced AI

AI designed to respect user privacy by limiting data access, ensuring confidentiality, and allowing individuals to control their data. Privacy-enhanced AI promotes autonomy and reduces risks of privacy intrusion.

Prohibited AI systems

AI systems whose applications are banned within the EU due to their capacity to harm fundamental rights, including systems using subliminal techniques, social scoring, or remote biometric identification in public spaces for surveillance purposes.

Provider’s Quality Management System

Internal processes established by providers to maintain compliance, quality, and safety standards throughout an AI system’s lifecycle, including during development, testing, and post-market monitoring.

Quantitative Bias Testing

A method for analyzing statistical biases in AI model outputs, ensuring they do not disproportionately affect specific groups. Quantitative bias testing supports fairness and compliance with anti-discrimination standards.

Quantitative Risk Assessment

A structured process quantifying potential risks posed by AI systems, evaluating the likelihood and severity of adverse impacts. This assessment informs risk mitigation strategies and promotes responsible AI deployment.

Real-Time AI Systems

AI systems capable of generating instant or near-instantaneous responses, often in high-stakes scenarios. These systems must be robust, secure, and reliable to mitigate risks associated with time-sensitive applications.

Regulatory Compliance

The requirement for AI systems to adhere to applicable EU regulations governing safety, transparency, and user protection. Regulatory compliance ensures AI systems meet legal standards for operation within the European market.

Remote Biometric Identification System

AI systems used to identify individuals from a distance, often in public spaces, by comparing live biometric data to reference databases, with stringent requirements for transparency, data protection, and authorization.

Responsible AI

The development and use of AI systems that prioritize ethical principles, transparency, and human rights. Responsible AI practices minimize harm, enhance accountability, and support compliance with regulatory standards.

Risk Assessment Frameworks

Structured methodologies for evaluating and mitigating risks associated with AI systems, ensuring that systems operate within acceptable risk levels and are equipped to handle potential adverse outcomes.

Risk Management System

A continuous, lifecycle-wide process to identify, assess, and mitigate risks posed by AI systems, ensuring their safe, ethical, and compliant operation within their intended use environments.

Risk Tolerance

The level of risk an organization or stakeholder is prepared to accept to achieve objectives. In AI, risk tolerance varies based on legal requirements, organizational priorities, and the specific context of system deployment.