AI Governance Glossary

The principle and practice of being responsible for AI system decisions and their consequences. Accountability mechanisms include clear attribution of responsibility, ability to explain decisions, and consequences for failures or harm.

An autonomous software entity that perceives its environment, makes decisions, and takes actions to achieve specific goals. AI agents can be simple (rule-based) or complex (machine learning-based).

A systematic, independent examination of an AI system to assess compliance with regulations, standards, and ethical principles. AI audits evaluate performance, bias, safety, security, and documentation.

The field of philosophy and practice focused on ensuring AI systems are developed and deployed responsibly, with consideration for fairness, transparency, accountability, and human values.

The frameworks, policies, and processes organizations implement to manage AI systems responsibly. Encompasses risk management, compliance, ethics, transparency, and human oversight throughout the AI lifecycle.

A systematic evaluation of how a proposed or deployed AI system might affect individuals, groups, and society. Examines potential risks including discrimination, privacy violations, and societal impacts.

The ability to understand, critically evaluate, and responsibly interact with AI systems. AI literacy encompasses technical knowledge, ethical awareness, and practical skills for working with AI.

A comprehensive set of policies, procedures, and tools for managing AI systems throughout their lifecycle. Includes governance, risk management, quality assurance, monitoring, and compliance activities (ISO 42001 standard).

The discipline and practice of ensuring AI systems operate reliably, securely, and without causing harm. Encompasses technical safety (robustness, security), operational safety (monitoring, controls), and societal safety (fairness, human oversight).

The confidence that stakeholders place in AI systems based on their understanding of system capabilities, limitations, transparency, and track record. Built through accountability, explainability, and consistent safe performance.

Systematic and unfair discrimination that occurs when AI systems produce consistently inaccurate or prejudicial outcomes for particular groups. Can result from biased training data, flawed model design, or misaligned objectives.

The practice of algorithmic systems producing different treatment or outcomes for individuals or groups based on protected characteristics (race, gender, age, etc.). Distinguished from statistical difference by intent and legal standards.

The degree to which the logic, data inputs, and decision-making processes of an algorithm are visible and understandable to stakeholders. Key component of trustworthy AI and regulatory compliance.

The process of identifying and measuring unfair or discriminatory patterns in AI system outputs. Uses statistical tests, fairness metrics, and data analysis to quantify bias across demographic groups.

Techniques and strategies to reduce or eliminate unfair discrimination in AI systems. Includes data balancing, algorithmic adjustments, threshold optimization, and fairness constraints during model development.

Formal recognition by CSOAI that an individual, organization, or AI system meets established standards for AI governance, safety, compliance, or performance. CSOAI certifications include CAGP, CASA, CEAAS, and CWA.

Adherence to applicable laws, regulations, standards, and organizational policies. In AI governance, compliance requires meeting requirements from frameworks like EU AI Act, NIST AI RMF, and ISO 42001.

Ongoing, real-time assessment of AI system performance, behavior, and compliance. Enables early detection of drift, bias drift, security threats, and regulatory violations through automated and manual analysis.

The formal process of determining whether an AI system complies with applicable requirements (regulatory, standard, organizational). Often includes documentation review, testing, audits, and certification by qualified assessors.

CSOAI's governance body and the conceptual foundation of the Council of Safety for AI. Represents the collective expertise and consensus-driven approach to AI governance across 33 specialized agent perspectives.

A distributed computing concept where a system continues to function correctly even if some nodes (agents) provide incorrect or malicious information. Used in CSOAI's 33-Agent Council for robust decision-making.

The frameworks and processes for managing data quality, access, privacy, and use throughout an organization. Critical for AI governance as data quality directly impacts AI system safety, fairness, and compliance.

Measures and practices to safeguard personal and sensitive data from unauthorized access, misuse, breaches, and other threats. Regulated by GDPR, CCPA, and similar privacy laws.

A subset of machine learning using neural networks with multiple layers to learn hierarchical representations of data. Powers large language models, computer vision systems, and other advanced AI applications.

The process of moving an AI model from development and testing into production where it serves real users or makes actual decisions. Requires thorough validation, monitoring setup, and ongoing governance.

CSOAI's expert body of AI safety professionals, ethicists, and regulators who review complex AI incidents, provide governance guidance, and establish best practices. Organizations can request DSRB reviews for critical systems.

The continuous monitoring process to identify when AI model performance, behavior, or predictions change over time (model drift), data characteristics change (data drift), or environmental conditions change (concept drift).

The comprehensive investigation and assessment of an AI system's risks, compliance status, performance characteristics, and governance practices. Required before deploying high-risk systems and merging organizations.

European Union regulation establishing comprehensive rules for AI systems based on risk levels. Defines prohibited practices, high-risk requirements, transparency rules, and conformity assessments. Enforcement timeline extends through 2026.

The quality of being understandable and interpretable. In AI context, explainability refers to the ability to provide clear, human-understandable explanations for why an AI system made specific decisions or recommendations.

The practice of incorporating ethical considerations and safeguards into AI systems from the initial design phase rather than attempting to add them later. Proactive approach to building trustworthy AI.

The principle and practice of ensuring AI systems treat all individuals and groups equitably without discrimination. Fairness metrics quantify whether outcomes are proportional across demographic groups.

Framework of principles for trustworthy AI: Fairness (equitable treatment), Explainability (understandable decisions), Accountability (clear responsibility), Transparency (open communication). Guides responsible AI development.

A large-scale AI model trained on diverse, broad data to serve as a base for multiple downstream applications. Examples include GPT models, BERT, and Llama. Foundation models demonstrate general AI capabilities across many tasks.

A structured set of guidelines, standards, and practices for implementing AI governance. Examples include NIST AI RMF, EU AI Act, and CSOAI's SOAI-PDCA framework. Frameworks provide methodologies and best practices.

General Data Protection Regulation - EU law governing personal data protection and privacy. Requires lawful basis for processing, transparency, data subject rights, and impact assessments. Critical for AI systems processing personal data.

AI systems with broad capabilities applicable across many different domains and use cases rather than specialized for single tasks. Foundation models and large language models are primary examples. Subject to specific EU AI Act requirements.

A comprehensive structure defining how an organization manages AI systems responsibly. Includes policies, procedures, accountability mechanisms, risk management, compliance monitoring, and human oversight structures.

AI systems that pose significant risks to fundamental rights, safety, or public interest. Under EU AI Act, high-risk systems (e.g., in law enforcement, employment, credit assessment) require conformity assessments, documentation, human oversight, and post-market monitoring.

The practice of maintaining human involvement in critical AI decisions and monitoring AI system behavior. Required for high-risk systems to ensure humans can understand, intervene in, or override AI decisions.

AI system design where humans remain actively involved in decision-making processes. Humans provide feedback, validate decisions, and intervene when necessary. Improves safety, fairness, and user acceptance of AI systems.

A systematic evaluation of potential consequences from deploying an AI system. Examines effects on individuals, groups, society, and environment across dimensions like fairness, privacy, employment, and fundamental rights.

International standard for AI management systems. Provides requirements for establishing, implementing, maintaining, and continuously improving AI governance. Includes risk management, quality assurance, monitoring, and documentation requirements.

The process of documenting, analyzing, and reporting problematic events involving AI systems (bias incidents, failures, security breaches, unintended consequences). Critical for continuous improvement and regulatory compliance.

Techniques used to circumvent safety measures and content filters in AI systems, often large language models. Users attempt to manipulate AI to generate harmful, inappropriate, or prohibited content.

Measurable value demonstrating effectiveness of AI systems and governance processes. AI governance KPIs include fairness metrics, model accuracy, compliance status, incident response time, and stakeholder satisfaction.

AI models trained on vast amounts of text data to predict and generate human language. Examples include GPT models and Claude. LLMs are foundation models with broad capabilities across language understanding and generation tasks.

Legal responsibility for harm or losses caused by AI systems. Liability regimes vary by jurisdiction and risk level; high-risk systems typically have stricter liability requirements. Enterprises need liability insurance and governance practices.

Subset of AI where systems learn patterns from data and improve performance without being explicitly programmed. Includes supervised learning, unsupervised learning, and reinforcement learning approaches.

Standardized documentation describing an AI model's performance, intended use, characteristics, and limitations. Includes training data, fairness metrics, performance across demographic groups, recommended use cases, and known limitations.

Processes and structures for managing AI models throughout their lifecycle: development, validation, deployment, monitoring, updates, and retirement. Ensures consistent quality, safety, and compliance across all models.

Continuous observation and analysis of deployed AI model performance, behavior, and health. Tracks metrics for accuracy, bias, latency, security, and identifies drift or anomalies requiring intervention.

CSOAI's founding pledge to prioritize human well-being and create opportunities for displaced workers. Ensures AI safety efforts directly translate into meaningful employment, training, and economic support.

National Institute of Standards and Technology AI Risk Management Framework. Voluntary, flexible framework for identifying, measuring, and managing AI risks. Provides guidance across design, development, deployment, and monitoring phases.

Legal and ethical principle ensuring AI systems do not unfairly distinguish between individuals or groups based on protected characteristics. Enforced through fairness requirements and bias testing.

Organization for Economic Cooperation and Development principles for responsible stewardship of trustworthy AI: human-centered values, transparency, robustness, accountability, and human oversight.

Risks arising from deficiencies in AI system operations, monitoring, maintenance, or response procedures. Includes model failures, drift, security incidents, and inadequate incident response.

Plan-Do-Check-Act continuous improvement methodology. Plan governance approach, Do implementation, Check results and compliance, Act on findings to improve. CSOAI's SOAI-PDCA framework applies PDCA to AI governance.

Ongoing surveillance and evaluation of AI systems after deployment to detect performance issues, safety failures, bias drift, security threats, and compliance violations. Required for high-risk systems under EU AI Act.

Approach of integrating data privacy and protection measures from the initial design phase of AI systems rather than adding them later. Minimizes data collection, implements anonymization, and respects user rights.

AI applications and techniques explicitly forbidden under regulations like EU AI Act. Examples include real-time facial recognition for mass surveillance, social scoring systems, and manipulation causing psychological harm.

CSOAI's economic support program for displaced workers. Provides training, income support, and employment opportunities in AI safety careers, operationalizing the Maternal Covenant commitment.

Systematic approach to ensuring AI systems and processes meet consistent standards throughout their lifecycle. Includes quality assurance, testing, validation, monitoring, and continuous improvement practices.

Adversarial testing practice where teams attempt to break, exploit, or manipulate AI systems to identify vulnerabilities, failure modes, and security issues before deployment.

Controlled environment where organizations can test innovative AI applications with relaxed regulatory requirements while maintaining safety oversight. Enables experimentation while protecting users.

Systematic process of identifying, analyzing, and evaluating potential risks from an AI system. Determines risk level (low, medium, high) and informs appropriate governance, monitoring, and compliance requirements.

Categorization of AI systems by their potential impact and risk level. EU AI Act defines prohibited, high-risk, limited-risk, and minimal-risk categories. NIST and ISO 42001 provide additional classification schemes.

The ability of AI systems to maintain safe and reliable performance under adverse conditions, unexpected inputs, or distribution shifts. Robust systems handle edge cases gracefully and resist adversarial attacks.

Condition of AI systems operating without causing harm. Encompasses technical safety (system reliability), operational safety (monitoring and controls), and societal safety (fairness, human oversight, accountability).

Isolated environment where AI systems or code are tested safely without affecting production systems or accessing sensitive data. Used for testing, development, and security evaluation.

CSOAI's proprietary AI governance methodology combining Safety-Oriented AI (SOAI) principles with Plan-Do-Check-Act (PDCA) continuous improvement cycles. Provides structured approach to implement responsible AI governance.

Any individual or group affected by or having interest in an AI system. Includes users, organizations, regulators, society, and potentially impacted communities. Stakeholder engagement is essential for responsible AI governance.

Established specifications, guidelines, and requirements that AI systems should meet. Examples include ISO 42001, NIST AI RMF, OECD Principles. Standards promote consistency, quality, and best practices.

Long-term viability and positive impact of AI systems. Addresses environmental impact (energy efficiency), social sustainability (equitable access), economic sustainability (job creation), and operational sustainability (maintainability).

China's standardization technical committee developing AI security and governance standards. Provides guidelines for algorithm transparency, data protection, and security assessments for high-impact AI systems.

Comprehensive records describing AI system architecture, training data, performance metrics, limitations, and use cases. Required for regulatory compliance, system understanding, and knowledge transfer.

The practice of making AI systems, their decisions, and their impacts visible and understandable to stakeholders. Includes explainability, documentation, and communication about system capabilities and limitations.

AI systems that are reliable, fair, transparent, accountable, and aligned with human values and societal norms. Built through ethics by design, robust governance, and continuous monitoring.

Under EU AI Act, AI practices posing unacceptable risks to human safety or rights are prohibited outright. Examples include social scoring, real-time facial recognition, and manipulation causing psychological harm.

Process of confirming that an AI system meets specified requirements and performs as intended before deployment. Includes functional testing, performance validation, bias testing, and safety verification.

Risks arising from using external AI systems, models, or services from third-party vendors. Includes security risks, performance risks, compliance risks, and dependency risks.

CSOAI's trained AI safety analysts who monitor, audit, and report on AI systems. Part of global community creating meaningful employment while protecting humanity from AI risks.

Technique of embedding imperceptible markers or patterns in AI-generated content to indicate its provenance and authenticity. Helps combat misinformation and deepfakes.

Field and set of techniques focused on making AI system decisions transparent and understandable to humans. Includes LIME, SHAP, attention mechanisms, and other interpretability methods.

Security principle of never assuming trust and continuously verifying all access and interactions. Applied to AI governance by continuously monitoring system behavior, data access, and integrity.