Trusted AI Glossary
A working vocabulary for leaders navigating AI governance, risk, and operations.
A
Adversarial Attack — A technique where malicious actors deliberately manipulate inputs to deceive an AI system into making incorrect predictions or decisions. These attacks exploit vulnerabilities in how models process data, potentially causing systems to misclassify images, bypass security controls, or produce harmful outputs.
AI Ethics Board — A cross-functional governance body responsible for establishing policies, reviewing high-risk AI applications, and providing guidance on ethical considerations. Organizations like IBM and Google have established these structures to centralize accountability and ensure consistent application of AI principles across the enterprise.
AI Literacy — The organizational capability to understand AI systems’ potential, limitations, and risks at all levels of the enterprise. The EU AI Act explicitly mandates AI literacy requirements, recognizing that effective governance depends on employees being equipped to work with and oversee AI systems responsibly.
AI Risk Management Framework (AI RMF) — A voluntary framework published by NIST in January 2023 that provides structured guidance for managing AI risks throughout the system lifecycle. It organizes risk management activities into four core functions (Govern, Map, Measure, Manage) and offers a flexible foundation that many organizations use to anchor their governance programs.
Algorithmic Accountability — The principle that organizations must be able to identify who is responsible for AI system outcomes and how decisions were made. This extends beyond technical explanations to encompass clear ownership structures, defined escalation paths, and mechanisms for affected individuals to seek recourse.
B
Bias (in AI) — Systematic errors in AI outputs that produce unfair outcomes for particular groups, often reflecting historical inequities in training data or flawed model design. Bias can manifest in hiring algorithms that disadvantage certain demographics, credit models that discriminate by geography, or healthcare systems that underserve specific populations.
C
Concept Drift — A change in the underlying relationship between input data and the target variable that a model was trained to predict. Unlike data drift, concept drift means the fundamental patterns have shifted—for example, customer behavior changing after a major economic event—requiring model retraining rather than just data adjustments.
Continuous Monitoring — The practice of systematically tracking AI system performance, behavior, and outputs in production environments on an ongoing basis. This operational discipline enables organizations to detect degradation, bias emergence, and anomalies before they cause significant harm, closing the gap between what AI systems should do and what they actually do.
D
Data Drift — A shift in the statistical properties of production data compared to the data used to train an AI model. When input data distributions change—due to seasonality, market shifts, or evolving user behavior—model performance typically degrades, making drift detection a critical component of production AI management.
Data Poisoning — An attack vector where adversaries deliberately corrupt training data to compromise model behavior, either by injecting malicious samples or manipulating existing data. The effects may remain dormant until triggered by specific inputs, making data provenance and integrity controls essential safeguards.
E
Explainability — The degree to which humans can understand how an AI system reaches its outputs, decisions, or recommendations. Explainability exists on a spectrum—from global explanations of overall model behavior to local explanations of individual predictions—and is essential for building stakeholder trust, enabling oversight, and meeting regulatory requirements.
EU AI Act — The European Union’s comprehensive legal framework for artificial intelligence, which entered into force in August 2024. It establishes a risk-based classification system with corresponding obligations for transparency, human oversight, and technical documentation, setting the global benchmark for AI regulation that affects any organization operating in EU markets.
F
Fairness (in AI) — The principle that AI systems should produce equitable outcomes across different demographic groups and avoid perpetuating or amplifying discrimination. Fairness is context-dependent and often involves trade-offs among competing definitions, such as equal outcomes, equal error rates, or equal treatment, which require deliberate choices aligned with organizational values.
G
Governance Framework — The integrated system of policies, structures, roles, and processes that guide how an organization develops, deploys, and manages AI systems. Effective frameworks address the entire AI lifecycle, establish clear accountability, and integrate with existing enterprise risk management rather than operating as a parallel compliance function.
Guardrails — Technical and procedural controls that constrain AI system behavior within acceptable boundaries. These may include content filters, output validation rules, human-in-the-loop checkpoints, or hard limits on system actions, providing defense-in-depth protection against harmful or unintended outcomes.
H
Human-in-the-Loop (HITL) — A design pattern where human judgment is integrated into AI system workflows, typically for high-stakes decisions, edge cases, or quality assurance. HITL approaches balance automation efficiency with human oversight, though effectiveness depends on ensuring humans have sufficient context, time, and authority to meaningfully intervene.
Hallucination — An AI system output that is fluent and confident but factually incorrect, fabricated, or nonsensical. Particularly prevalent in large language models, hallucinations pose significant risks in enterprise contexts where users may not recognize false information—making output validation and grounding techniques essential safeguards.
I
ISO/IEC 42001 — The first international standard for AI management systems, specifying requirements for establishing, implementing, and continually improving AI governance within organizations. It provides a certifiable framework that integrates with other management system standards, offering organizations a structured approach to demonstrating governance maturity to stakeholders.
M
Model Card — A standardized documentation format that describes an AI model’s intended use, performance characteristics, limitations, and ethical considerations. Originally proposed by researchers at Google, model cards support transparency and informed decision-making by providing stakeholders with essential information to evaluate whether a model is appropriate for their context.
Model Drift — The degradation of AI model performance over time as the relationship between inputs and real-world outcomes changes. Model drift is inevitable in dynamic environments and manifests as declining accuracy, increasing errors, or shifting bias patterns—making ongoing monitoring and periodic retraining essential operational practices.
O
Observability (AI) — The technical capability to understand the internal state and behavior of AI systems based on their external outputs, logs, and metrics. In production AI contexts, observability platforms provide visibility into model performance, data quality, prediction distributions, and failure modes, enabling teams to diagnose issues and maintain system health.
R
Red Teaming (AI) — A structured practice where dedicated teams attempt to find vulnerabilities, failure modes, and harmful behaviors in AI systems before deployment. Borrowed from cybersecurity, AI red teaming involves systematic adversarial testing, including prompt injection, jailbreaking attempts, and edge case exploration, to identify risks that conventional testing may miss.
Responsible AI — An umbrella term for the practices, principles, and organizational commitments that ensure AI systems are developed and deployed in ways that are ethical, fair, transparent, and aligned with human values. While related to trusted AI, responsible AI often emphasizes the ethical dimensions, whereas trusted AI encompasses the full governance and operational infrastructure.
T
Transparency — The practice of openly communicating how AI systems work, what data they use, and how they influence decisions affecting individuals. Transparency operates at multiple levels, from public disclosure of AI use to technical documentation for auditors, and is increasingly mandated by regulations like the EU AI Act for high-risk applications.
Trust Layer — An architectural component that provides governance controls, safety mechanisms, and compliance safeguards for AI systems. Salesforce’s Einstein Trust Layer exemplifies this approach, embedding capabilities like data masking, toxicity detection, and audit logging directly into the platform to protect enterprise AI deployments.
This glossary will expand as Trusted AI evolves. Have a term you’d like defined? Message me.