AI Risk & Governance Glossary
A definitive, searchable resource for AI governance and compliance terminology.
Browse & Search Terms
A/B Testing
An experimental method comparing two versions of a system (A and B) on a random subset of users to determine which performs better.
Accountability
The principle that individuals and organizations can be held responsible for decisions made by AI systems they deploy, and must be able to explain those decisions to affected parties.
Adversarial Attack
An attempt to manipulate or deceive an AI model by deliberately crafting inputs designed to cause misclassification or unintended behaviour.
AI Safety
The field of research and practice focused on ensuring AI systems behave as intended, avoid unintended harms, and remain aligned with human values.
Algorithmic Fairness
The principle that machine learning systems should treat individuals or groups equitably, not systematically disadvantaging people based on protected characteristics.
Audit Trail
A complete, timestamped record of actions taken by an AI system and the decisions made, enabling post-hoc review and accountability.
Confusion Matrix
A table showing predicted vs. actual values for a classification task, enabling calculation of metrics like precision, recall, and false positive/negative rates.
Consent and Transparency
The principles that individuals affected by AI systems should be informed about their use, understand how they work, and have choice in whether to participate.
Context Window
The maximum length of text (measured in tokens) that a language model can consider at once when generating responses.
Control Objective
A specific, measurable statement of what a control is meant to achieve against a named risk. It describes the outcome, not the mechanism — and it is only useful if a failure to meet it would be detectable.
Data Poisoning
An attack in which an adversary deliberately injects false, misleading, or malicious data into a training dataset to compromise model performance or introduce hidden vulnerabilities.
Data Quality
The degree to which data is accurate, complete, consistent, and representative of the phenomena it purports to measure.
Differential Privacy
A mathematical framework ensuring that algorithms protect individual privacy by adding carefully calibrated noise, making it impossible to infer whether a specific individual's data was used.
Embedding
A numerical representation of text, images, or other data in a high-dimensional space where similar items are positioned close together.
Emergent Capabilities
Abilities that appear in AI models at certain scales but were not explicitly trained for, suggesting complex behaviors arising from scale alone.
EU AI Act
European Union legislation that classifies AI systems by risk level and imposes requirements on high-risk applications, including impact assessments, documentation, and human oversight.
Explainability
The ability to give a person an account of why a system produced a particular output, in terms they can act on. Explainability is a property of the account and its audience, not of the model — an opaque model can be explainable, and an interpretable one can fail to be explained.
Fairness & Bias
The principle and practice of ensuring AI systems do not produce systematically different outcomes for individuals or groups based on protected characteristics or vulnerable attributes.
Federated Learning
A machine learning approach where models are trained across decentralized data sources without centralizing sensitive data in one location.
Fine-Tuning
The process of adapting a pre-trained model to a specific task or domain by training it on a smaller, task-specific dataset.
Governance Framework
A structured set of policies, roles, and controls designed to guide organizational decision-making and oversight of AI systems.
Hallucination
Output that is fluent, confident and false. It arises from the same mechanism that produces correct output — a model predicting plausible continuations — which is why it is best understood as a property of the technology to be bounded rather than a defect to be fixed.
Human-in-the-Loop (HITL)
A design pattern where human judgment and decision-making are retained in the process, with AI providing recommendations or automating parts of a workflow but not full autonomy.
Impact Assessment
A systematic evaluation of the potential positive and negative effects of deploying an AI system on individuals, communities and society. Distinct from a data protection impact assessment, which examines the lawfulness of the data processing rather than the consequences of the decision.
Incident Response
A structured process for detecting, assessing, containing, and resolving adverse events involving AI systems, such as model failure, security breach, or fairness harm.
ISO/IEC 42001
International standard for AI management systems, establishing requirements for organizations to manage risks and opportunities associated with AI systems.
Jailbreak
A technique for inducing a language model to produce output that its safety training was intended to refuse. It targets the model's alignment rather than any particular application, and works substantially the same way wherever that model is deployed.
Large Language Model (LLM)
A deep learning model trained on vast amounts of text data, capable of understanding and generating human-like language across diverse topics.
Machine Learning
A subset of artificial intelligence where systems learn patterns from data without being explicitly programmed for every scenario.
Model Card
A document detailing the capabilities, limitations, intended use cases, performance metrics, and known risks of a machine learning model.
Model Collapse
A phenomenon where training language models on synthetic data generated by other language models leads to degradation of model quality and diversity over successive generations.
Model Deployment
The process of putting a trained AI model into production where it makes real-world predictions or decisions affecting users or business outcomes.
Model Drift
A change in the relationship between a model's inputs and the environment it was fitted to, such that the model's assumptions no longer hold. The model itself is unchanged; the world it was trained on has moved.
Model Interpretability
An intrinsic property of a model: the degree to which a person can understand the mechanism by which it turns inputs into outputs. A linear model or a shallow decision tree is interpretable by construction. A deep network is not, however good an explanation you can construct for it afterwards.
Model Monitoring
Continuous observation of a deployed model inputs, outputs, performance metrics, and behaviour to detect degradation, drift, misuse, or security anomalies.
Model Versioning
The practice of recording which model, trained how and on what, was deployed when — enabling comparison, rollback and reproducibility. For systems built on models the organisation did not train, it extends to versioning the surrounding assembly, since that is the only part it controls.
Multimodal AI
AI systems that process and integrate multiple types of data (text, images, audio, video) to understand and respond to the world.
Neural Network
A computational model inspired by biological neurons, consisting of interconnected layers of artificial nodes that learn to recognize patterns.
NIST AI Risk Management Framework
A U.S. framework providing guidance on managing risks in AI systems across four functions: Map, Measure, Manage, and Govern.
Performance Metric
A quantitative measure of how well an AI model or system performs on a task, such as accuracy, precision, recall, F1 score, or domain-specific measures.
Performance Regression
A decline in model performance over time, typically caused by data distribution shifts, concept drift, or changes in the environment.
Prompt Engineering
The practice of crafting and refining prompts to guide language models toward desired outputs, treating the prompt itself as a design problem.
Prompt Injection
An attack in which untrusted content reaching a language model is interpreted as instructions rather than data, causing the model to act on an attacker's intent instead of the operator's. It is a property of how the application assembles its context, not a flaw in any particular model.
Retrieval-Augmented Generation (RAG)
A technique where an LLM retrieves relevant documents or data before generating a response, grounding outputs in actual information rather than relying on training data alone.
Risk Appetite
The level and types of risk an organisation is willing to accept in pursuit of its strategic objectives. Applied to AI, a single organisational posture is rarely enough — appetite has to be set separately across distinct risk dimensions and tiers of decision consequence.
Synthetic Data
Artificially generated data created by algorithms or models, used as a substitute for or supplement to real data in training and testing.
Testing AI Systems
Systematic evaluation of AI models and systems to verify they perform as intended, identify failure modes, and assess risks.
Third-Party Risk
The exposure an organisation carries through dependencies on external parties who build, train, host or operate the AI systems it uses. In AI the exposure is characteristically deeper than the contract: beneath most vendors sits a model provider, and beneath that an infrastructure provider.
Token
A unit of text that a language model processes, typically a word, subword, or character. Models operate on sequences of tokens, not raw text.
Training Data
The dataset used to train a machine learning model, from which the model learns patterns and relationships.
Transformer
A deep learning architecture that uses self-attention mechanisms to process sequences of data, foundational to modern large language models.
Transparency
The disclosure of information about how an AI system is built, what it does, what it is used for, and what risks it poses, communicated to relevant stakeholders.