137 AI Terms, Plain and Simple
What is this AI glossary for?
This A–Z guide explains 137 terms used across modern artificial intelligence, machine learning, generative AI and automation. Use it for a quick definition, then follow the related reading links where a term connects to the eBook catalogue.
How do these AI terms connect in practice?
Machine learning is the broad pattern-learning approach; large language models are one family of models built for language; generative AI describes systems that create new content. In a business workflow, those models may be paired with human-in-the-loop review, watched for model drift, and checked for algorithmic bias.
Example: a bank may use anomaly detection for fraud, natural language processing for document review, and explainability plus human approval for high-impact decisions. A manufacturer may combine computer vision, predictive analytics and a digital twin to spot defects and test changes before touching a production line.
For topic-led explanations, continue to AI topic guides, comparisons, or the eBook catalogue.
A
- AI Agent
- A software system that can interpret a goal, plan steps, use tools, take actions and check results with limited human direction.
- AI Assistant
- A conversational interface that helps a person find information, create content, make decisions or complete tasks.
- AI Governance
- The policies, roles, controls and evidence used to ensure AI systems are lawful, safe, accountable and aligned with organisational objectives.
- AI Literacy
- The practical ability to understand what AI can and cannot do, use it responsibly and judge its outputs critically.
- AI Safety
- The discipline of reducing the chance that AI systems cause harm through errors, misuse or unintended behaviour.
- AI Search
- Search that uses natural-language understanding, semantic matching and often generative answers rather than exact keywords alone.
- Algorithm
- A defined set of steps or rules that a computer follows to solve a problem or produce an output.
- Algorithmic Bias
- Systematic unfairness in an automated system, often introduced through data, labels, objectives, design choices or deployment context. Related reading: The House Always Knows: AI, Gambling, and the Ethics of Personalized Gaming.
- Alignment
- The effort to make an AI system behave consistently with human intentions, values, policies and safety constraints.
- Annotation
- Adding labels or explanatory information to data so it can be used to train or evaluate a model.
- Anomaly Detection
- Identifying observations or behaviours that differ sharply from an expected pattern. Related reading: Artificial Intelligence for Cyber Security: A Practical Guide to Data Breach Prevention.
- Artificial General Intelligence (AGI)
- A hypothetical AI system able to learn and perform a very wide range of intellectual tasks at a broadly human level.
- Artificial Intelligence (AI)
- The field of building computer systems that perform tasks associated with intelligence, including perception, language, prediction, reasoning and decision-making. Related reading: The Artificial Intelligence Revolution: From Algorithms to Consciousness.
- Attention Mechanism
- A model component that assigns greater weight to the parts of an input most relevant to the current prediction.
- Automatic Speech Recognition (ASR)
- Technology that converts spoken language into written text.
- Automation
- Using technology to complete tasks or processes with reduced manual effort.
- Autonomous System
- A system that senses its environment, makes decisions and acts with little immediate human control.
B
- Benchmark
- A standardised test or dataset used to compare AI capability, quality, speed, safety or cost.
- Big Data
- Data collections whose scale, speed or variety makes them difficult to manage with traditional tools alone.
- Black-Box Model
- A model whose internal reasoning or decision process is difficult for people to inspect or explain.
C
- Chain of Thought
- Intermediate reasoning steps used by some models to work through a task; these are not always exposed to users.
- Chatbot
- A text or voice interface that responds to user messages, ranging from scripted menus to advanced conversational AI.
- Classification
- Assigning an input to one or more predefined categories.
- Clustering
- Grouping similar data points together without predefined labels.
- Computer Vision
- AI techniques that interpret images and video, including object detection, recognition and tracking. Related reading: AI in Aviation: Transforming Safety and Sustainability and Artificial Intelligence in Construction: Building a Sustainable Future.
- Concept Drift
- A change in the relationship between inputs and outcomes that makes a deployed model less accurate over time.
- Confabulation
- A confident-sounding but unsupported model output.
- Context Window
- The maximum amount of information a model can consider in a single request or conversation.
- Conversational AI
- AI designed to understand and generate natural dialogue across text or speech.
- Copilot
- An AI assistant embedded in a tool or workflow to support a person rather than operate fully independently.
D
- Data Augmentation
- Expanding a training dataset by creating modified or synthetic examples.
- Data Drift
- A change in live input data compared with the data used to train or validate a model.
- Data Governance
- The rules and responsibilities used to manage data quality, access, security, lineage and lawful use.
- Data Leakage
- When training or evaluation includes information unavailable in real-world use, producing misleading results. Related reading: Artificial Intelligence in Industry: A Comprehensive Guide.
- Deep Learning
- Machine learning based on neural networks with many processing layers.
- Demand Forecasting
- Using historical and current data to estimate future demand. Related reading: Artificial Intelligence in Logistics: Optimizing Efficiency and Sustainability and AI-Powered Smart Grid: Revolutionizing Electricity Distribution and Generation.
- Deterministic System
- A system that produces the same output whenever it receives the same input under the same conditions.
- Diffusion Model
- A generative model that creates data, especially images, by reversing a process that gradually adds noise.
- Digital Twin
- A data-linked virtual representation of a real object, process or environment used for monitoring and simulation. Related reading: Smart Buildings: AI-Powered Efficiency and Sustainability and AI-Powered Smart Grid: Revolutionizing Electricity Distribution and Generation.
- Distillation
- Training a smaller model to imitate a larger one so it can run with lower cost or latency.
E
- Edge AI
- Running AI models on local devices or near the source of data.
- Embedding
- A numerical representation that captures meaning or similarity so data can be compared mathematically.
- Enterprise AI
- The governed use of AI across an organisation to improve operations, decisions, products or services.
- Evaluation
- The structured process of testing an AI system for quality, safety, reliability, fairness, cost and suitability.
- Existential Risk (X-Risk)
- A debated risk category in which advanced AI could cause irreversible civilisation-wide harm.
- Explainability
- Methods for describing why a model produced a particular prediction, recommendation or decision. Related reading: Artificial Intelligence and the Law: Case Studies and Future Trends.
- Explainable AI (XAI)
- Methods and practices used to make AI behaviour and decisions understandable.
F
- Fairness
- The aim of preventing unjustified differences in how an AI system treats people or groups.
- Federated Learning
- Training a shared model across separate devices or organisations without centralising raw data. Related reading: Artificial Intelligence for Cyber Security: A Practical Guide to Data Breach Prevention.
- Few-Shot Learning
- Guiding a model to perform a task by providing a small number of examples.
- Fine-Tuning
- Further training a pre-trained model so it performs better for a particular domain, style or task.
- Foundation Model
- A large, broadly trained model that can be adapted for many downstream applications.
- Fraud Detection
- Using rules, statistics or machine learning to identify suspicious transactions or behaviour. Related reading: Artificial Intelligence in Banking: Revolutionizing Finance and Data Security.
G
- Generative Adversarial Network (GAN)
- A generative architecture in which one network creates examples while another judges whether they look real.
- Generative AI
- AI that creates new text, images, audio, video, software code or other content. Related reading: The AI Music Revolution: Creativity, Controversy, and Collaboration and Lights, Camera, Algorithm: AI’s Role in Modern Filmmaking.
- Generative Pre-Trained Transformer (GPT)
- A transformer-based language model trained on large datasets to generate or interpret language.
- Grounding
- Connecting an AI response to trusted data, tools, rules or real-world context.
- Guardrail
- A technical or procedural control that restricts unsafe, irrelevant or non-compliant AI behaviour.
H
- Hallucination
- An AI output that sounds plausible but is false, unsupported or inconsistent with available evidence.
- Human-in-the-Loop
- A workflow in which a person reviews, approves, corrects or overrides AI output. Related reading: Digital Diagnosis: How AI is Revolutionizing Healthcare.
- Human-on-the-Loop
- A workflow in which AI acts independently while a person supervises and can intervene.
- Hyperparameter
- A setting that influences how a model learns, such as learning rate or batch size.
I
- Inference
- Using a trained model to generate a prediction, classification, recommendation or piece of content.
- Inference Cost
- The computing or service cost incurred each time a trained model processes a request.
- Instruction Tuning
- Training a model on instructions and desired responses so it follows requests more reliably.
- Interpretability
- How readily people can understand a model's internal logic, structure or behaviour.
K
- Knowledge Graph
- A structured network of entities and relationships that helps systems connect facts.
L
- Large Language Model (LLM)
- A language model trained on very large text collections to understand and generate natural language. Related reading: The Dumbening: How AI is Reshaping Our Minds.
- Latency
- The delay between sending a request to an AI system and receiving its response.
- LoRA (Low-Rank Adaptation)
- A fine-tuning method that trains small added matrices instead of changing every model weight.
M
- Machine Learning (ML)
- A branch of AI in which models learn patterns from data rather than being programmed for every case. Related reading: Artificial Intelligence in Industry: A Comprehensive Guide.
- Machine Learning Operations (MLOps)
- The practices and tools used to deploy, monitor, govern and maintain machine learning systems.
- Model
- A mathematical or computational system that turns inputs into predictions, decisions or generated outputs.
- Model Card
- A document describing a model's intended use, limitations, evaluation results and risks.
- Model Context Protocol (MCP)
- An open protocol for connecting AI applications to tools and data sources through a consistent interface.
- Model Drift
- A decline in model performance as data, behaviour or real-world conditions change. Related reading: Artificial Intelligence in Banking: Revolutionizing Finance and Data Security.
- Model Registry
- A controlled catalogue storing model versions, metadata, approvals and deployment status.
- Multi-Agent System
- A system in which several AI agents collaborate, specialise or check one another.
- Multimodal Model
- A model that can work with more than one type of data, such as text, images, audio or video.
N
- Natural Language Generation (NLG)
- Using software to produce human-readable or spoken language from data or other inputs.
- Natural Language Processing (NLP)
- AI techniques for analysing, understanding, transforming and generating human language. Related reading: From Reporters to Robots: How AI is Reshaping Journalism.
- Natural Language Understanding (NLU)
- The part of language AI concerned with extracting meaning, intent, entities and context.
- Neural Network
- A machine learning architecture made of connected processing units arranged in layers.
- No-Code AI
- Tools that allow people to configure or use AI applications without conventional software coding.
O
- Object Detection
- Identifying both the category and location of objects within an image or video.
- Open-Source Model
- A model whose code, weights or documentation are available under terms permitting inspection or reuse.
- Optimisation
- Adjusting a model, prompt or system to improve accuracy, cost, latency or safety.
- Overfitting
- When a model learns its training data too closely and performs poorly on new examples.
P
- Parameter
- A numerical value learned during training that influences model output.
- Parameter-Efficient Fine-Tuning (PEFT)
- Methods that adapt a model by training only a small portion of its parameters.
- Personalisation
- Tailoring content, recommendations or experiences using context, preferences or behaviour. Related reading: The AI Behind Your Feed: Personalization, Moderation, and the Future of Social Media.
- Pre-Training
- Training a model on a broad dataset before adapting it for specific tasks.
- Precision
- Among the items a model marked positive, the proportion that were actually positive.
- Predictive Analytics
- Using historical data and statistical or machine learning techniques to estimate future outcomes. Related reading: Artificial Intelligence Revolution in Manufacturing: Modernizing Operations, Maintenance, and Service Delivery.
- Predictive Maintenance
- Using operational and sensor data to predict equipment failure before it occurs. Related reading: The Autonomous Revolution: Artificial Intelligence and the Future of the Automotive Industry and AI Revolution in Railways: Modernizing Travel for a Smarter Future.
- Prompt
- The instruction, context or input supplied to a generative AI system.
- Prompt Engineering
- Designing and testing prompts, context and examples to produce more reliable AI outputs.
- Prompt Injection
- An attack in which hostile instructions attempt to override an AI system's rules.
- Provenance
- Evidence showing where data, content, outputs or decisions came from.
Q
- Quality Gate
- A defined check that output must pass before publication or production.
- Quantisation
- Reducing numerical precision so a model uses less memory and runs faster.
R
- Reasoning Model
- A model optimised to spend more computation analysing a problem before answering.
- Recall
- Among all items actually positive, the proportion correctly identified by a model.
- Recommendation System
- A system that ranks or suggests products, content or actions.
- Reinforcement Learning
- Training through interaction, using rewards and penalties. Related reading: Game AI Unleashed: From Finite State Machines to Machine Learning.
- Reinforcement Learning from Human Feedback (RLHF)
- Using human preferences to train a model towards more helpful behaviour.
- Responsible AI
- Designing, deploying and governing AI with attention to safety, fairness, privacy and accountability.
- Retrieval-Augmented Generation (RAG)
- Retrieving relevant external information and supplying it to a generative model before it answers.
S
- Semantic Search
- Search based on meaning and context rather than exact words alone.
- Sentiment Analysis
- Classifying the attitude or emotion expressed in content.
- Small Language Model (SLM)
- A language model designed to require less memory, computing power and cost.
- Speech-to-Text
- Converting spoken audio into written text.
- Steerability
- How effectively a model's behaviour can be guided by instructions, examples or controls.
- Structured Data
- Information organised according to a defined schema.
- Summarisation
- Producing a shorter representation that preserves the most important information.
- Supervised Learning
- Training a model using examples paired with known answers or labels.
- Synthetic Data
- Artificially generated data designed to resemble real data.
T
- Temperature
- A generation setting that influences how predictable or varied model output is.
- Text-to-Speech (TTS)
- Converting written text into generated spoken audio.
- Token
- A unit of text processed by a language model.
- Tokenisation
- Breaking text into smaller units that a model can process.
- Tool Use
- The ability of an AI model or agent to call external functions, software or databases.
- Training Data
- The examples used to teach a machine learning model.
- Transformer
- A neural-network architecture built around attention mechanisms.
- Transparency
- Providing understandable information about how an AI system was built, tested and used.
U
- Underfitting
- When a model is too simple or insufficiently trained to capture important patterns.
- Unstructured Data
- Information without a fixed tabular format, such as documents, images and recordings.
- Unsupervised Learning
- Training methods that discover patterns without labelled answers.
V
- Validation Set
- Data kept separate from training and used to tune choices or compare models.
- Vector Database
- A database designed to store and search numerical embeddings by similarity.
- Vision-Language Model (VLM)
- A multimodal model trained to connect visual content with language.
- Voice Synthesis
- Generating artificial speech that imitates natural human voice characteristics.
W
- Weak AI
- AI designed for a limited task or domain rather than general intelligence.
- Weights
- The learned numerical values inside a neural network.
- Workflow Automation
- Coordinating tasks, systems and decisions so a process progresses automatically.
Z
- Zero-Shot Learning
- Performing a task without labelled examples of that exact task.
- Zero-Trust AI
- Applying zero-trust security principles so every user, tool, model action and data request is verified.
Want the full context?
Use the glossary for the quick definition, then explore the 40-book library for practical context, case studies and deeper analysis.