AI Governance and Machine Learning Glossary

Comprehensive vocabulary flashcards covering fundamental concepts, algorithms, frameworks, and governance principles in artificial intelligence and machine learning.

Accountability (AI Governance)

The obligation and responsibility of developers, deployers, and distributors of an AI system to ensure the system operates ethically, fairly, transparently, and compliant with applicable rules and regulations, ensuring actions, decisions, and outcomes can be traced back to the entity responsible for it.

Standard neural networks

often called a Feedforward Neural Network or a Multi-Layer Perceptron (MLP)—is a basic machine learning model inspired by the human brain that maps inputs to outputs through connected layers of mathematical units called neurons

Convolutional neural networks (CNNs)

deep learning algorithms designed to process grid-structured data—most commonly images, video, and audio. They mimic how the human visual cortex works to recognize patterns, edges, shapes, and entire objects without requiring manual feature engineering

Unique characteristics of AI that require a comprehensive approach to governance

Complexity

Opacity

Potential for harm or misuse

Speed And scale

Data dependency

Autonomy

Probabilistic vs deterministic output

Internet of Things

A vast network of physical objects—known as "smart devices"—embedded with sensors, software, and network connectivity that allow them to collect, share, and act on data in real time.

Accountability (Privacy)

The implementation of appropriate technical and organizational measures to ensure and demonstrate that personal data handling complies with relevant laws (such as the EU GDPR and APEC's Cross Border Privacy Rules), rooted in the fair information practices principle of due diligence.

Accuracy

The degree to which an AI system correctly performs its intended task, measuring its performance and effectiveness in producing correct outputs based on input data.

Active Learning

A subfield of AI and machine learning (also called query learning) in which an algorithm selects some of the data it learns from, actively requesting additional data points to learn better rather than learning passively from all provided data.

Adaptive Learning

An educational method that adjusts and tailors instruction and content to the specific needs, abilities, and learning pace of individual students to provide an optimized, personalized learning experience.

Adversarial Attack

A safety and security risk to an AI model instigated by manipulating the model (such as introducing deceptive or malicious input data) to cause malfunctions, incorrect outputs, or unsafe behaviors.

Agentic AI

AI systems (also known as AI agents) designed to autonomously pursue complex goals, make decisions, and take actions with limited human oversight or supervision.

AI Assurance

A combination of frameworks, policies, processes, and controls that measure, evaluate, and promote safe, reliable, and trustworthy AI through conformity assessments, audits, certifications, and standards compliance.

AI Audit

A systematic review of an AI system or portfolio of AI systems to verify that it operates as intended and complies with relevant laws, regulations, and standards, helping map unforeseen risks during impact assessment and development.

AI Governance

A system of organizational-level frameworks, practices, and processes that help stakeholders implement, oversee, and manage the use and risks of AI to ensure ethical, responsible, and compliant deployment.

Algorithm

A computational procedure or defined set of instructions and rules designed to perform a specific task, solve a problem, or build a machine learning model.

Artificial General Intelligence (AGI)

Theoretical AI systems (also known as strong AI) possessing human-level intelligence, general cognitive abilities, and strong generalization capability to reason, learn, adapt, and solve problems across a broad range of domains without specialized training.

Artificial Intelligence (AI)

A broad field of computer science and engineered systems dedicated to simulating intelligent behavior and using computational techniques to automate tasks previously performed by humans.

Artificial Narrow Intelligence (ANI)

AI systems (also called narrow AI or weak AI) engineered to perform specific, narrowly defined tasks (such as Siri or Alexa) that excel within their domain but lack the ability to transfer knowledge to other areas.

Artificial Super Intelligence (ASI)

Hypothetical AI systems that surpass human intelligence across virtually every domain, including scientific capability, creativity, general wisdom, and social skills.

Broad Artificial Intelligence

AI systems that can perform a broader set of tasks than narrow AI (often combining multiple systems, such as autonomous vehicles or agentic swarms) but lack the complete generalized, human-like capabilities of AGI.

Automated Decision-Making

The process of reaching a decision strictly through technological means without human involvement.

Autonomy

The ability of an AI system to operate independently without human intervention.

Computational Bias

A systematic error or deviation from the true value of a prediction that originates directly from a machine learning model's assumptions or the training data itself.

Cognitive Bias

Inaccurate individual human judgment or distorted thinking that can permeate and influence an AI model or system.

Societal Bias

Systemic prejudice, favoritism, or discrimination against or in favor of an individual or group that can permeate an AI system through mechanisms such as selection bias in training data.

Blockchain

A shared, immutable digital ledger of transactions copied across a network of nodes that verifies transactions via consensus and permanently locks them together in sequence using cryptographic hashes.

Bootstrap Aggregating (Bagging)

An ensemble machine learning technique that trains multiple versions of a model on random subsets of a dataset and aggregates their predictions to increase model stability and accuracy.

Chatbot

An AI application designed to simulate human conversation and interaction, typically utilizing natural language processing and large language models to understand and reply to text or other media.

Classification Model

A supervised machine learning model (also known as a classifier) trained on pre-labeled data to map incoming inputs into predefined discrete categories or classes.

Clustering

An unsupervised machine learning method that evaluates patterns and groups unlabeled data points together into clusters based on shared similarities without predefined category labels.

Compute

The processing resources available to a computer system, including hardware components like CPUs and GPUs, required for memory, storage, running applications, and training models.

Computer Vision

A subfield of AI focused on enabling computers to process, analyze, and interpret visual inputs such as digital images and videos.

Conformity Assessment

A formal evaluation, often executed by an entity independent of the developer, to verify that an AI system satisfies mandated requirements such as data governance, risk management, and cybersecurity.

Contestability

The principle ensuring that AI decisions, outputs, and actions can be challenged, reviewed, or redressed by human beings.

Corpus

A large collection of structured or unstructured texts or data utilized by a computer to extract patterns, generate outcomes, or train machine learning algorithms.

Counterfactual

A hypothetical alternative scenario evaluated by modifying one or more input variables to determine what change would alter an AI model's output or decision.

Data Drift

A shift that happens over time when the operational input data changes so that it no longer statistically resembles the data used during model training, leading to performance degradation.

Data Leak

The accidental exposure of sensitive, confidential, proprietary, or personal information resulting from flawed security controls, human error, or storage misconfigurations.

Data Poisoning

An adversarial attack where malicious data is intentionally injected into a training dataset to manipulate the learning process and corrupt model performance.

Data Provenance

The documented history and audit trail that tracks records across their full lifecycle—including origins, transformations, actors, and custody—to establish data integrity and transparency.

Data Quality

A measure of how well a dataset satisfies requirements for its intended use, characterized by accuracy, completeness, validity, consistency, timeliness, and fitness for purpose.

Decision Tree

A supervised machine learning model that maps decisions and their possible consequences in a hierarchical, branching tree structure.

Deep Learning

A subfield of AI and machine learning that employs multi-layered artificial neural networks with hidden layers to process complex raw inputs such as imagery, audio, and text.

Deepfakes

Audio or visual content that has been altered, generated, or synthesized using generative AI to produce realistic yet fabricated depictions.

Diffusion Model

A class of generative models that synthesizes realistic media by iteratively removing noise from a randomized signal guided by a prompt.

Discriminative Model

A machine learning model designed to identify boundaries between data points to classify, distinguish, or map input features to discrete target categories.

Disinformation

Manipulated, synthetic, or false audio, visual, or textual content created and distributed with the deliberate intent to deceive or cause harm.

Entropy

A mathematical measure of randomness, unpredictability, or disorder in a dataset; higher entropy corresponds to greater uncertainty in predictive outcomes.

Expert System

A domain-specific AI system that applies inference rules to an established knowledge base to replicate human expert decision-making.

Explainability

The ability to provide post-hoc, human-understandable justifications or explanations (using tools like SHAP or LIME) for how a complex, black-box model reached a specific output. Related to black box systems.

Interpretability

The degree to which a human can directly inspect, follow, and comprehend the internal logic, mathematical structure, and behavior of a model from the ground up (inherent to transparent, white-box models).

Exploratory Data Analysis (EDA)

Initial data discovery techniques performed prior to model training to uncover patterns, identify anomalies, detect outliers, and understand relationships between variables.

Fail-Safe Plans

Predefined operational contingency procedures designed to execute if an AI system displays unintended, erratic, or dangerous behaviors.

Fairness (AI Governance)

The principle that an AI system must treat individuals and groups with equitable consistency, ensuring decisions do not produce direct or disparate negative impacts based on sensitive attributes like race, gender, or religion.

Federated Learning

A distributed machine learning technique where models are trained locally across multiple edge devices or servers, and only parameter updates (not raw user data) are sent to a central server to aggregate into a global model.

Fine-Tuning

The supervised process of adapting a broad pre-trained foundation model to a specialized task by training it further on a smaller, labeled dataset.

Foundation Model

A large-scale, general-purpose model trained on vast and diverse datasets (also called frontier AI) that exhibits versatile capabilities and serves as the baseline architecture for downstream tasks.

Generalization

The capacity of a machine learning model to accurately apply the patterns, structures, and relationships learned during training to previously unseen data.

Generative AI

A branch of AI that models the underlying probability distributions of training data to synthesize original content such as text, imagery, audio, synthetic code, and video.

Greedy Algorithm

An algorithm that makes the locally optimal choice at each individual decision step with the aim of finding an immediate solution, without considering the overall long-term optimal outcome.

Ground Truth

The empirically verified, objective reality or verified target labels of a dataset used as a standard benchmark to validate the accuracy of an AI model.

Hallucination

A phenomenon (also called confabulation) where a generative AI model outputs convincing, grammatically plausible claims that are factually incorrect or unsupported by source information.

Human in the Loop (HITL)

A design model that integrates human oversight, interaction, or validation directly into an AI system's automated decision-making workflow.

Human-Centric AI

An approach to artificial intelligence design, development, and governance that explicitly prioritizes human well-being, rights, values, and autonomy, augmenting human capabilities rather than replacing them.

Impact Assessment

A formal evaluation process (also termed a risk assessment) conducted to identify, analyze, and mitigate potential legal, societal, ethical, and economic ramifications of deploying an AI system.

Inference

The functional phase where a trained machine learning model ingests real-world input data to generate predictions, scores, or decisions.

Large Language Model (LLM)

A deep learning model with high parameter capacity trained on massive text corpora to understand, generate, and process natural language tasks based on learned statistical word sequences.

Machine Learning

A subfield of AI that uses algorithms to discover patterns within training data, iteratively improving performance on tasks without explicit rule-based programming.

Machine Learning Model

A learned mathematical representation of underlying patterns and relationships extracted from training data via an algorithm, capable of performing inference on new inputs.

Misinformation

Inaccurate or false information distributed without deliberate malicious intent or intent to cause harm.

Model Card

A standardized, short transparency document disclosing vital information about an AI model, such as intended use cases, performance benchmarks, and evaluation across diverse demographic subgroups.

Model Context Protocol (MCP)

An open standard and communication protocol acting as a universal adapter that enables AI models to securely interface with external tools, local files, and data sources.

Multimodal Model

A machine learning model designed to simultaneously process, integrate, or produce multiple distinct forms of input or output data, such as combining text, audio, and images.

Natural Language Processing (NLP)

A subfield of AI that gives computers the ability to read, interpret, analyze, translate, and generate human language in text or speech formats.

Neural Network

A computational model inspired by biological nervous systems featuring interconnected nodes organized into input, hidden, and output layers to model complex nonlinear relationships.

Overfitting

An error in machine learning where a model learns the training data and noise too closely, failing to generalize to new, unseen data.

Parameters

Internal variables (such as weights in neural networks) that a model learns and adjusts from training data to define how inputs are mapped to predictions.

Post-Processing

Techniques and mathematical adjustments applied to the raw outputs of an AI model after inference to calibrate results, boost fairness, or satisfy business logic.

Preprocessing

Data preparation steps executed prior to model training—such as imputation of missing values, normalization, feature extraction, and categorical encoding—to enhance data quality and model reliability.

Prompt Engineering

The deliberate and iterative craft of designing, structuring, and refining prompts to guide generative AI models toward accurate and desirable outputs.

Random Forest

An ensemble supervised learning algorithm that constructs multiple decision trees using bootstrap aggregation and random feature subsets to produce robust and stable predictions. Helpful for datasets with missing values or high complexity.

Red Teaming

A structured security evaluation process where an independent team simulates adversarial attacks, jailbreaks, and prompt injections to uncover model flaws, biases, and vulnerabilities before public deployment.

Reinforcement Learning

A machine learning paradigm where an autonomous agent learns to maximize cumulative rewards through trial-and-error interactions and environmental feedback in the form of penalties and rewards.

Reinforcement Learning with Human Feedback (RLHF)

A fine-tuning methodology that incorporates human evaluations of model outputs to train a reward model, aligning AI behaviors and values with human intent.

Retrieval-Augmented Generation (RAG)

An architecture that enhances a generative language model by retrieving relevant authoritative passages from external databases before generating an answer.

Robustness

An attribute of an AI system denoting its resilience to withstand perturbations, adversarial tampering, and environmental shifts while maintaining operational accuracy.

Semi-Supervised Learning

A machine learning approach that trains an algorithm on a large quantity of unlabeled data alongside a small subset of labeled data to reduce labeling costs.

Small Language Models (SLMs)

Compact language models with reduced parameter counts and smaller training data footprints, optimized for fast inference, local edge deployment, and lower resource requirements.

Supervised Learning

A foundational type of machine learning in which models are trained on input-output pairs consisting of features and known target labels to perform classification or regression.

Synthetic Data

Artificially generated data that mathematically mirrors the properties, patterns, and statistical distribution of real-world data, used when actual data is scarce, restricted, or sensitive.

System Card

A transparency document that details how multiple individual models and supporting components interoperate within an overarching multi-agent or system network.

Transfer Learning

A machine learning approach where a model trained on one foundational task transfers its learned representations to accelerate training and improve accuracy on a related target task.

Transformer Model

A neural network architecture that learns context and maintains relationships between sequence data, such as words in a sentence. It does so by using attention, i.e., focusing on the most important and relevant parts of the input sequence. This helps to improve model accuracy. For example, in language learning tasks, by attending to surrounding words, the model can understand a word's meaning in the context of the whole sentence.

Trustworthy AI

An umbrella term describing AI systems developed and governed according to foundational principles such as fairness, privacy, safety, accountability, transparency, and explainability.

Turing Test

A benchmark proposed by Alan Turing evaluating whether a machine can exhibit conversational behavior that is indistinguishable from that of a human judge.

Underfitting

A scenario where a machine learning model is overly simplistic or insufficiently trained, leaving it unable to capture the underlying structure of the data and resulting in poor performance.

Unsupervised Learning

A machine learning approach where an algorithm evaluates unlabeled data to uncover underlying relationships, structures, and clusters without external guidance or targets.

Validation Data

A distinct subset of data used during development to evaluate intermediate model performance, fine-tune hyperparameters, and prevent overfitting before final testing.

Variance

A statistical metric measuring the degree of dispersion of data points around their mean; in machine learning, excessive variance leads to models fitting training noise (overfitting).