Governance breaks down when people use the same word to mean different things. One team calls a system "high-risk" because it touches customers; a lawyer means something precise under Annex III of the EU AI Act. An engineer says a model is "explainable" because it prints feature weights; an auditor wants a documented method a regulator will accept. A shared vocabulary is not pedantry. It is what lets legal, risk, data science, security, and the board make the same decision from the same facts, and it is what turns a policy document into controls people can actually apply.
This glossary defines the terms that show up in AI governance programs, regulations, and vendor pitches in 2026. It is grouped by theme rather than strictly A to Z, because related ideas are easier to learn together. Each entry is short and current, with a pointer to a related term where that helps. Bookmark it, share it with a new committee member, and use it to settle the next debate about what a word actually means.
Core concepts
The foundational ideas every governance conversation rests on. Get these straight and the rest of the vocabulary falls into place.
| Term | Definition |
|---|---|
| AI governance | The system of policies, roles, processes, and controls that keeps AI systems legal, accountable, and aligned with what the organization intends, across their whole lifecycle. It connects AI to the same oversight enterprises already apply to finance and security. |
| Responsible AI | An umbrella practice for building and running AI that is fair, transparent, safe, and accountable. It is often the internal program name; governance is the machinery that makes it real. |
| AI ethics | The study of the moral principles that should guide how AI is designed and used: fairness, autonomy, human dignity, and the avoidance of harm. It informs policy but is not itself an enforcement mechanism. |
| AI safety | The field concerned with preventing AI systems from causing unintended harm, covering robustness, alignment with human intent, and control of highly capable models. Distinct from security, which defends against deliberate attackers. |
| Algorithmic accountability | The principle that a named person or body must answer for an algorithm's decisions and their effects, with the ability to explain, contest, and correct them. Without a clear owner, accountability is a slogan. |
| Transparency | How openly a system's purpose, data, logic, and limits are disclosed to the people it affects and to those who oversee it. Under EU AI Act Article 50, some systems must tell users they are dealing with AI. |
| Explainability (XAI) | The degree to which the reasons for a model's output can be understood by a human. See SHAP and LIME under technical terms for the common methods. |
| Fairness | The property that a model does not produce systematically worse outcomes for protected or vulnerable groups. Several formal definitions exist and they can conflict, so teams choose metrics deliberately rather than assuming one "fair." |
| Bias | Systematic error that skews outputs, arising from data, model design, or how a system is deployed. Not all bias is unlawful, but discriminatory bias creates legal and reputational exposure. See bias audit. |
| Trustworthy AI | A framing used by the EU and OECD for AI that is lawful, ethically sound, and technically reliable. NIST AI RMF describes the trustworthy characteristics an organization should aim for, such as validity, safety, and accountability. |
| AI literacy | The knowledge and skills that let staff use and oversee AI appropriately for their role. EU AI Act Article 4 has required providers and deployers to ensure a sufficient level of AI literacy since 2 February 2025. |
Regulations and standards
The laws and voluntary frameworks that now define what "compliant" means. Most enterprise programs map one control set to several of these at once.
| Term | Definition |
|---|---|
| EU AI Act | The first comprehensive horizontal AI law, Regulation (EU) 2024/1689, sorting systems into four tiers: unacceptable, high, limited, and minimal risk. Obligations phase in from 2025 through 2027, with a few embedded high-risk cases to 2 August 2028. |
| Prohibited practices | Uses banned outright under EU AI Act Article 5, including social scoring, untargeted scraping of facial images, and manipulative or exploitative systems. In force since 2 February 2025. |
| High-risk AI system | Systems in the Annex III use cases (such as biometrics, employment, credit, and essential services) or safety components under Annex I. They carry the strictest obligations; most Annex III duties apply from 2 August 2026, with some deferred to 2 December 2027. |
| GPAI (general-purpose AI) | Models with broad capability across many tasks, such as large language models. Providers face transparency and technical documentation duties under the EU AI Act from 2 August 2025. |
| Systemic-risk model | A GPAI model presumed to pose systemic risk when trained above 10^25 floating-point operations. It triggers extra duties: model evaluation, red-teaming, and serious-incident reporting. |
| GPAI Code of Practice | A voluntary code finalized in July 2025 that gives GPAI providers a practical way to show compliance with the Act's model obligations. |
| NIST AI RMF | A voluntary US framework organized around four functions: Govern, Map, Measure, and Manage. Its Generative AI Profile (NIST AI 600-1, 2024) adds twelve GenAI-specific risk categories. |
| ISO/IEC 42001 | The 2023 standard for an AI management system, and the first that an organization can be certified against by an accredited body. See conformity assessment. |
| ISO/IEC 23894 | A 2023 guidance standard for managing AI-specific risk, aligned with the general enterprise risk vocabulary of ISO 31000. |
| ISO/IEC 42005 | A 2025 guidance standard for conducting AI system impact assessments in a structured, repeatable way. It is a companion to ISO/IEC 42001. |
| ISO/IEC 42006 | A 2025 standard setting requirements for the bodies that audit and certify AI management systems, which underpins the credibility of a 42001 certificate. |
| Conformity assessment | The process of demonstrating that a high-risk AI system meets legal requirements before it reaches the market, done either by self-assessment or through a notified body under the EU AI Act. |
| US state AI laws | With no federal statute, states set the pace. Texas TRAIGA took effect 1 January 2026 and the Colorado AI Act takes effect 30 June 2026; New York City Local Law 144 requires bias audits for automated hiring tools. |
Risk and assurance
The vocabulary of finding, sizing, controlling, and proving. Much of it is borrowed from decades of practice in banking and audit, then adapted to AI.
| Term | Definition |
|---|---|
| AI risk taxonomy | A structured catalog of the ways an AI system can fail or cause harm, used to keep risk assessment consistent across teams. The NIST and MIT AI risk repositories are widely cited starting points. |
| Risk register | A living record of identified AI risks with owners, severity ratings, mitigations, and status. It is the day-to-day backbone of a risk program, not a one-time document. |
| Impact assessment (AIA / FRIA) | A structured evaluation of a system's effects before deployment. EU AI Act Article 27 requires a fundamental rights impact assessment (FRIA) for certain deployers of high-risk systems; ISO/IEC 42005 gives a general method. |
| DPIA | A data protection impact assessment, required under GDPR Article 35 when processing is likely to result in high risk to individuals. It often runs alongside an AI impact assessment. |
| Model risk management (MRM) | The discipline of identifying, measuring, and controlling the risk that a model is wrong or misused, across its lifecycle. Long established in banking and now extended to AI. |
| SR 11-7 | US Federal Reserve and OCC supervisory guidance from 2011 on model risk management. Many firms treat it as the reference point they extend to cover AI and machine-learning models. |
| Three lines of defense | A governance model where the business owns and manages risk (first line), risk and compliance functions oversee and challenge it (second line), and internal audit gives independent assurance (third line). |
| Red-teaming | Structured adversarial testing that probes a model for harmful, biased, or unsafe outputs before and after release. It is a required practice for systemic-risk GPAI under the EU AI Act. |
| AI audit | An independent examination of an AI system or program against a standard, regulation, or internal policy, producing documented evidence of where it does and does not comply. |
| Bias audit | A specific test for disparate outcomes across demographic groups. New York City Local Law 144 mandates an independent bias audit for automated employment decision tools before they are used. |
| Audit trail | A tamper-evident record of the actions, decisions, and changes across a system's lifecycle. It is what lets you reconstruct what happened, and who approved it, when a regulator or customer asks. |
A definition is not a control. Naming a system "high-risk" changes nothing until that label triggers an impact assessment, an owner, monitoring, and an audit trail. The value of shared terms is that they route work to the right process automatically.
Technical and model terms
The artifacts and mechanisms governance teams review, even when they do not build the models themselves. Knowing what these are is the difference between oversight and rubber-stamping.
| Term | Definition |
|---|---|
| AI lifecycle | The stages a system moves through: problem framing, data collection, development, validation, deployment, monitoring, and retirement. Governance attaches controls at each stage rather than only at launch. |
| Model card | A short structured document describing a model's intended use, performance, evaluation data, and known limitations. The format was proposed by Google researchers in 2019 and is now common practice. |
| Datasheet | Short for "datasheets for datasets": a record of a dataset's motivation, composition, collection method, and known biases, so downstream teams understand what they are training on. |
| System card | A document describing a whole AI system, including several models and their safeguards, used by some frontier developers to explain overall behavior and risk mitigations. Broader than a single model card. |
| SHAP | SHapley Additive exPlanations. A method that assigns each input feature a contribution to a specific prediction, grounded in cooperative game theory. A widely used explainability technique. |
| LIME | Local Interpretable Model-agnostic Explanations. It approximates a complex model near a single prediction with a simple, readable model to explain that one case. |
| Model drift | The decay of a model's performance over time as live data diverges from training data. It splits into data drift (inputs shift) and concept drift (the relationship being modeled shifts). See monitoring. |
| Monitoring | Continuous tracking of a deployed model's inputs, outputs, and performance to catch drift, errors, and misuse. It is the operational heart of post-market oversight, not an optional add-on. |
| Foundation model | A large model trained on broad data that can be adapted to many downstream tasks. The GPT, Claude, and Gemini families are examples. |
| LLM | A large language model: a foundation model trained to predict and generate text. LLMs underpin most current generative AI applications. |
| RAG | Retrieval-augmented generation. It grounds an LLM's answers in documents retrieved at query time, improving accuracy and providing sources, which reduces hallucination. |
| Hallucination | A confident but false or fabricated model output. It is a core reliability risk for LLMs and a named category in the NIST Generative AI Profile. |
| Provenance and watermarking | Techniques for marking and tracing the origin of AI-generated content. EU AI Act Article 50 requires providers to mark synthetic content in a machine-readable way, applying from 2 August 2026. |
| Data governance | The policies and controls over how data is sourced, quality-checked, secured, and used. It is the foundation trustworthy AI is built on; weak data governance undermines every model above it. |
| Data lineage | The documented path of data from its origin through every transformation to its use in a model. It is essential for auditability, debugging, and answering where a training example came from. |
Roles and operating model
Who decides, who reviews, who answers, and the records that make ownership real. Governance fails most often on unclear roles, not missing policies.
| Term | Definition |
|---|---|
| Chief AI Officer (CAIO) | The senior executive accountable for AI strategy, adoption, and governance. US federal agencies were directed to designate one under Office of Management and Budget guidance, and the title is now common in large enterprises. |
| AI governance committee | A cross-functional body (legal, risk, security, data, and business leaders) that sets policy, reviews high-risk use cases, and owns escalation. It is where trade-offs get decided, not just discussed. |
| AI ethics board | A group, sometimes including external members, that reviews the harder value questions a use case raises. In most organizations it advises rather than holds veto power. |
| RACI | A responsibility matrix marking who is Responsible, Accountable, Consulted, and Informed for each governance task. It is a simple tool for removing the ownership gaps that stall programs. |
| Human-in-the-loop (HITL) | A design where a person reviews or approves each AI decision before it takes effect. Appropriate for high-stakes decisions; contrast with human-on-the-loop. |
| Human-on-the-loop | A design where a person supervises AI that runs autonomously and can step in, rather than approving every individual action. It scales better but demands good monitoring. |
| Human oversight | The broader requirement, set out in EU AI Act Article 14, that high-risk systems be built so people can understand their output, monitor operation, and override them when needed. |
| AI inventory (registry) | A complete, maintained record of the AI systems and models an organization builds or buys. It is the single source of truth a governance program depends on; you cannot govern what you cannot see. |
| Use-case registry | A record of the business uses of AI, each tied to an owner, a risk classification, and an approval status. It complements the model inventory by capturing intent, not just technology. |
Emerging: agentic, GenAI, shadow AI
The fastest-moving corner of the vocabulary. These terms barely existed in mainstream governance a few years ago and now dominate the roadmap.
Generative AI (GenAI)
AI that produces new content such as text, images, code, or audio, typically built on foundation models. Its accessibility is what pushed AI into every department, and with it, shadow AI.
Agentic AI
AI that pursues goals over multiple steps with limited human input, planning actions and calling tools as it goes. Gartner expects agents embedded in a large share of enterprise software by the end of 2026, which is reshaping what governance has to cover.
AI agent
A software entity that perceives, decides, and acts toward a goal, often invoking tools, APIs, or other agents. Governing agents adds three questions a static model never raised: what identity it acts under, what permissions it holds, and how its actions are observed and traced.
Shadow AI
AI tools used inside an organization without approval or oversight. Surveys point to a majority of employees using unapproved AI tools, which is why automated discovery, rather than a voluntary survey, is how mature programs find them.
AI TRiSM
Gartner's term for AI trust, risk, and security management: the technical controls that turn governance policy into runtime reality through continuous monitoring, validation, and enforcement. In 2026 Gartner extended the concept to guardian agents that police other agents.
Guardrails
Automated controls that constrain a model's inputs and outputs, blocking unsafe, off-policy, or non-compliant behavior as it happens. They are a runtime complement to lifecycle governance, not a substitute for it.
Prompt injection
An attack that hides malicious instructions in text a model reads, hijacking its behavior or exfiltrating data. The Open Worldwide Application Security Project (OWASP) ranks it the top security risk for LLM applications.
Model Context Protocol (MCP)
An open standard for connecting AI models and agents to external tools and data sources. It makes integration easier and, in doing so, raises fresh permission, logging, and audit questions that governance now has to answer.
Why this section grows fastest. Agentic systems act, not just predict. That collapses the gap between a model's output and a real-world consequence, so the terms here, identity, permissions, guardrails, and observability, are becoming the center of gravity for governance programs, not a fringe.
Keeping the glossary operational
Definitions age. A glossary earns its place only when the terms connect to living records and real controls.
Three habits keep shared vocabulary useful rather than decorative. First, tie each term to an owner and a process: "high-risk" should trigger an impact assessment and a named approver, not just a label in a slide. Second, revisit the fast-moving entries, especially the agentic and GenAI ones, on a fixed cadence, because regulators and standards bodies are still writing these definitions in real time. Third, make the words match reality by grounding them in an inventory of the AI you actually run.
- Map each term to the framework clauses it satisfies, so one definition serves the EU AI Act, NIST AI RMF, and ISO 42001 at once.
- Attach terms to records: a "model card" entry should link to the model cards you hold, not just define the concept.
- Review quarterly and version the glossary, since a term like "systemic-risk model" carries a specific legal threshold that can change.
From policy to practice. Spreadsheets and ticket queues rarely keep up with how fast AI spreads across an enterprise, which is how shared terms drift back into meaning different things to different teams. Dedicated AI governance platforms turn this vocabulary into a live record of your actual AI, giving governance teams one place to discover, assess, monitor, and evidence every model and agent against frameworks like the EU AI Act, NIST AI RMF, and ISO 42001.
Key Takeaways
- Shared, precise terms are what let legal, risk, data, and security act on the same facts. Ambiguity is where governance quietly fails.
- Group terms by theme, not alphabet: core concepts, regulations, risk and assurance, technical, roles, and the fast-moving emerging set.
- Legal terms carry exact meanings. "High-risk," "GPAI," and "systemic-risk model" are defined in the EU AI Act, not open to interpretation.
- A definition is only useful when it routes to a control: an owner, an assessment, monitoring, and an audit trail.
- An AI inventory turns the whole glossary into a live record of the systems and agents you actually run.
Frequently asked questions
Responsible AI is the goal: AI that is fair, transparent, safe, and accountable. AI governance is the machinery that gets you there, the policies, roles, processes, and controls. In practice, responsible AI is often the program name and governance is what makes it operate.
They are related but not identical. Interpretability is how inherently understandable a model's internal mechanics are, for example a small decision tree. Explainability is the broader ability to give humans reasons for an output, including methods like SHAP and LIME applied to models that are not interpretable on their own.
Several. "Prohibited practices," "high-risk AI system," "GPAI," and "systemic-risk model" are defined in the EU AI Act with specific criteria and thresholds, such as the 10^25 FLOP mark for systemic risk. "DPIA" is defined by GDPR Article 35. Treat these as terms of art and check the source text before relying on them.
An AI inventory records the systems and models themselves, the technology you build or buy. A use-case registry records the business uses, each with an owner, a risk classification, and an approval status. You need both: one tracks what exists, the other tracks intent and accountability.
Because agents act rather than only predict. That introduces questions a static model never raised: what identity an agent operates under, what permissions it holds, how its steps are observed, and how it is stopped. Terms like agent identity, guardrails, and observability exist to make those controls governable.
Review it at least quarterly, and version it. The core concepts are stable, but the regulatory and emerging entries change quickly: standards are still being published (ISO/IEC 42005 and 42006 both arrived in 2025) and EU AI Act obligations phase in through 2027. A stale definition can send work to the wrong process.
