AI Glossary: AI terms explained with diagrams

AI Glossary

Plain-language definitions of the AI terms that come up in boardrooms and engineering stand-ups. Each entry has a diagram, an example, the reason it matters and the concept it is most often confused with.

176 terms

01Foundations16

Artificial IntelligenceAI

The field of building computer systems that perform tasks normally requiring human intelligence, such as understanding language, recognising images, making decisions and solving problems.

Machine LearningML

The branch of AI in which a system learns patterns from examples and uses them to make predictions or decisions, without a person writing the rules by hand.

Deep Learning

A branch of machine learning that uses neural networks with many layers to learn directly from raw data such as images, sound and text.

Neural Network

Artificial neural network

A computing system made of layers of simple connected units, loosely modelled on neurons, that learns by adjusting the strength of the connections between them.

Generative AIGenAI

AI that produces new content, such as text, images, audio, video or code, in response to a request, drawing on patterns learned from very large amounts of existing material.

Large Language ModelLLM

An AI model trained on very large amounts of text to predict the next token, which enables it to write, summarise, translate, analyse, answer questions and produce code.

Small Language ModelSLM

A language model with far fewer parameters than the largest models, built to run cheaply and quickly, often on a single device or on a company's own servers.

Foundation Model

A large AI model trained on broad data that can be adapted to many different tasks, and so serves as the common base on which applications are built.

Frontier Model

Frontier AI

One of the most capable AI models available at a given moment, at the leading edge of what the technology can do.

Open-weight Model

Open weights

An AI model whose trained weights are published, so that anyone can download it, run it on their own hardware and adapt it, within the terms of its licence.

Transformer

The neural network architecture behind nearly all modern language models, built around attention, a mechanism that lets the model weigh every part of the input against every other part at once.

Diffusion Model

A generative model that creates images, video or audio by starting from random noise and removing the noise step by step until a result that matches the request appears.

Multimodal AI

Multimodal model

AI that can take in, and sometimes produce, more than one kind of content, such as text, images, audio and video, within a single model.

Natural Language ProcessingNLP

The field of AI concerned with enabling computers to understand, interpret and produce human language, in text and in speech.

Computer Vision

The field of AI that enables computers to extract information from images and video: what is shown, where it is, and what is happening.

Artificial General IntelligenceAGI

A hypothetical AI system able to perform any intellectual task a human can, at a human level or above, across all domains and not only in one specialised area.

02How models work18

Token

The unit of text a language model reads and writes: a word, a piece of a word, or a punctuation mark.

Tokenization

The step that converts text into the sequence of tokens a model works on, and converts the model's output tokens back into text.

Context Window

The maximum amount of text, measured in tokens, that a model can take into account at one time: instructions, conversation, documents and its own reply together.

Parameters

Weights

The numbers inside a neural network that are adjusted during training and that together determine everything the model does.

Training

The process in which a model's parameters are adjusted, example by example, until its predictions match the data it is shown.

Pre-training

The first and largest stage of training, in which a model learns general patterns of language and knowledge from a vast body of text before being adapted to any task.

Training Data

The text, code, images and other material a model learns from during training; its content and quality shape what the model knows and how it behaves.

Inference

Running a trained model to produce an output: every answer, summary or tool call a model generates is an act of inference.

Next-token Prediction

The single operation at the core of a language model: given the text so far, estimate how likely each possible next token is.

Attention

The mechanism that lets a model weigh, for each token, which other tokens in the text matter most for understanding it.

Embedding

A list of numbers that represents the meaning of a piece of text, so that texts with similar meaning have similar numbers.

Latent Space

The internal, many-dimensional space in which a model represents what it has learned; positions and directions in it correspond to concepts and relationships.

Temperature

A setting that controls how much randomness a model uses when choosing each next token: low values give consistent output, high values give varied output.

Sampling

Top-p · top-k

The rules by which a model picks the next token from its probability list; top-k and top-p limit the choice to the most likely candidates.

Hallucination

Output that is fluent and confident but false or unsupported: an invented fact, citation, quote or detail that the model presents as real.

Knowledge Cutoff

The date after which a model has no information from its training; anything that happened later is unknown to it unless supplied at the time of use.

Reasoning Model

A language model that works through a problem step by step before answering, spending extra computation on planning, trying approaches and checking its own work.

Extended Thinking

A mode in which a model is given a budget of extra tokens to reason before it answers, with the budget set by the user or developer.

03Prompting & context16

Prompt

The input given to a model: the instructions, questions, examples and material it receives and responds to.

System Prompt

Instructions placed at the start of a conversation by the developer of an AI product, which set the model's role, rules, tone and the tools it may use.

Prompt Engineering

The practice of designing, testing and refining the instructions given to a model so that it performs a task reliably.

Context Engineering

The discipline of deciding what information is in a model's context window at each step: instructions, documents, tool results, history and memory.

Zero-shot Prompting

Asking a model to perform a task with instructions only, without showing it any examples of the desired output.

Few-shot Prompting

Including a small number of worked examples in the prompt so the model can infer the pattern, format and tone you want.

Chain of ThoughtCoT

A prompting technique in which the model is asked to work through a problem step by step before giving its final answer.

Prompt Chaining

Splitting a task into a fixed sequence of prompts, where the output of each step becomes the input of the next.

Meta-prompting

Using a model to write, critique or improve prompts for a model, in place of writing every prompt by hand.

Persona

A role, character or voice assigned to a model, usually in the system prompt, that shapes how it speaks and what perspective it takes.

Structured Output

Model output that follows a predefined format, typically JSON matching a schema, so that software can read and act on it without guesswork.

Grounding

Tying a model's answer to specific source material, such as documents, data or search results, so that its claims can be traced and checked.

Prompt Caching

A provider feature that stores the processed form of a repeated prompt prefix so that later requests reusing it are cheaper and faster.

Context Rot

The decline in a model's accuracy and focus as its context window fills with more tokens, even when the limit has not been reached.

Compaction

Replacing a long conversation history with a shorter summary so that an AI session can continue within its context window.

Memory

Any mechanism that lets an AI system carry information from one conversation or session to another, since the model itself retains nothing between requests.

04Knowledge & retrieval13

Retrieval-Augmented GenerationRAG

A technique in which relevant documents are retrieved at question time and placed in the model's context, so that it answers from your information and not only from its training.

Retrieval

The step that finds and fetches the pieces of information most relevant to a question from a larger body of content, so they can be given to a model.

Vector Database

A database built to store embeddings and to find, very quickly, the stored items whose meaning is closest to a query.

Semantic Search

Search that matches on meaning instead of exact words, so a query finds relevant content even when it uses different vocabulary.

Hybrid Search

Search that runs keyword matching and semantic matching together and merges their results into one ranked list.

Chunking

Splitting documents into smaller passages before indexing, so that each passage can be found and handed to a model on its own.

Reranking

A second pass that re-orders retrieved passages by how well each one answers the query, so that the best few are the ones the model receives.

Knowledge Base

An organised, maintained collection of an organisation's information, such as policies, procedures, product details and answers, that people and AI systems consult as the authoritative source.

Knowledge Graph

A way of storing knowledge as a network of entities, such as people, companies and products, and the named relationships between them, so that connections can be followed and queried.

GraphRAG

Graph RAG

A form of RAG that first builds a graph of the entities and relationships in a document collection, then uses it to answer questions that depend on connections or on the collection as a whole.

Agentic RAG

RAG in which an AI agent controls the retrieval: it decides what to search for and where, judges whether the results are enough, and searches again before it answers.

Citations

Source attribution

References attached to an AI answer that show which source each claim came from, so that a reader can check it.

Enterprise Search

Workplace search

Search across an organisation's own systems, such as documents, email, chat, wikis and business applications, from one place and limited to what each user is permitted to see.

05AI search & visibility11

Generative Engine OptimizationGEO

The practice of shaping content and online presence so that AI systems such as ChatGPT, Perplexity and Google's AI answers retrieve it, cite it and describe the brand accurately.

AI OptimizationAIO

An umbrella term for making a brand's content easy for AI systems to find, understand and represent correctly; the abbreviation is also widely used for Google's AI Overviews.

Answer Engine OptimizationAEO

Structuring content so that systems which answer questions directly, such as AI assistants, voice assistants and search answer boxes, select it as the answer.

AI Overviews

Google's AI-generated summaries that appear at the top of the search results page and answer the query directly, with links to the sources they draw on.

Answer Engine

A system that responds to a question with a written answer, usually backed by sources it has searched and read, in place of a list of links.

llms.txt

A proposed convention: a Markdown file at a website's root that gives AI models a short, curated guide to the site's most important content.

AI Crawler

AI bot

An automated program run by an AI company that visits web pages to collect their content, whether to train models, to build an AI search index or to answer a user's question live.

Query Fan-out

The technique in which an AI search system splits one question into several related searches, runs them at the same time and combines the results into a single answer.

Zero-click Search

A search that ends on the results page because the answer is shown there, so the user never clicks through to a website.

AI Visibility

AI search visibility

How often and how accurately a brand, product or website appears in the answers that AI assistants and AI search engines give.

Structured Data

Schema.org markup · schema markup

Standardised labels added to a web page's code that tell machines exactly what the content is: a product and its price, an article and its author, a question and its answer.

06Model customization10

Fine-tuning

Continuing the training of an already trained model on a smaller set of your own examples, so that its weights shift towards a particular task, style or output format.

LoRA

Low-Rank Adaptation

A fine-tuning method that leaves a model's original weights frozen and trains a small set of additional weights, called an adapter, which is stored as a separate file and applied on top of the model.

Transfer Learning

Reusing what a model learned on one task or dataset as the starting point for a different task, so that far less data, time and computing power are needed than when training from scratch.

Distillation

Knowledge distillation

Training a small model, the student, to reproduce the behaviour of a large one, the teacher, so that most of the quality is kept at a fraction of the cost and response time.

Quantization

Quantisation

Reducing the number of bits used to store each of a model's weights, for example from 16 to 4, so that the model needs less memory and runs faster, at the price of a small loss in quality.

Synthetic Data

Data that is generated artificially, by a model, a simulation or a set of rules, to stand in for or add to data collected from the real world.

Data Labeling

Data annotation · data labelling

Attaching the correct answer to each example in a dataset, such as a category, a tag, a box around an object or a quality score, so that a model can learn from it or be tested against it.

Supervised and Unsupervised Learning

The two classic ways a model learns from data: supervised learning uses examples that come with the correct answer, while unsupervised learning looks for structure in data that has no answers attached.

Reinforcement LearningRL

A way of training in which a system learns by acting: it tries actions, receives a reward or a penalty for the results, and gradually adopts the behaviour that earns the most reward over time.

Reinforcement Learning from Human FeedbackRLHF

A training technique in which people rate or compare a model's answers and the model is then adjusted to produce the kind of answers people preferred.

07Agents27

Agent

AI agent

An AI system that pursues a goal by deciding its own steps: it calls tools, looks at the results, and keeps going until the job is done.

Agentic AI

AI that works toward a goal across many steps with limited supervision: it plans, uses tools, checks its own results and adjusts.

Agentic Workflow

A process in which models carry out the steps, but the steps and their order are laid out in advance by people in code.

Agent Loop

Agentic loop

The repeating cycle at the core of every AI agent: the model reads the situation and chooses an action, a tool carries it out, the result is added to the context, and the cycle runs again until a stopping rule ends it.

ReAct

Reason + Act

A pattern for building agents in which a language model alternates between reasoning about what to do next, acting through a tool and reading the result, until it can answer.

Tool Use

Function calling

The mechanism that lets a model ask your software to run a function (search, look up a record, send a message) and then use the result.

Model Context ProtocolMCP

An open standard for connecting AI applications to external tools and data, so one connector works with any compatible assistant.

Agent2AgentA2A

An open protocol that lets AI agents built by different vendors on different frameworks discover each other, exchange tasks and report results.

Agent Skills

Reusable packages of instructions, scripts and reference files that an agent loads only when a task calls for them.

Agent Harness

Agent scaffold

The software around a language model that turns it into a working agent: it runs the loop, executes tool calls, manages context and memory, enforces permissions and decides when to stop.

Orchestration

The coordinating layer that decides which model, tool, agent or person does what, in which order, and keeps track of the state of the work.

Multi-agent System

A setup in which several AI agents, each with its own role, instructions and context, work together on one task.

Subagent

An agent started by another agent to handle one piece of a task in its own separate context, returning only the result.

Handoff

Handover

The transfer of responsibility for a task, together with the context needed to continue it, from one AI agent to another agent or to a person.

Planning

Task decomposition

An agent's ability to break a goal into ordered steps before acting, and to revise those steps as results come in.

Agent Memory

The ways an agent keeps track of its own work outside the context window: notes and files during a task, and stores of past runs, facts and procedures that it consults in later ones.

Long-running Agent

Long-horizon agent

An AI agent that works on one task for hours or days, across many context windows and sessions, with little or no human input along the way.

Computer Use

A capability that lets a model operate software the way a person does: it looks at the screen, then clicks, types and scrolls.

Browser Agent

Web agent · browsing agent

An AI agent that carries out tasks on websites by operating a web browser: it opens pages, reads them, clicks, fills in forms and moves between sites until the task is done.

Coding Agent

AI coding agent

An AI agent that carries out software tasks across a whole codebase: it reads the code, edits files, runs commands and tests, and delivers a change for a person to review.

Vibe Coding

Building software by describing what you want to an AI and accepting the code it writes without reading it, steering only by whether the result works.

Deep Research

Research mode

A feature of AI assistants that works on a question for several minutes: it plans, runs many searches, reads the sources it finds and writes a structured report with citations.

Copilot

AI copilot · AI assistant

An AI assistant built into a working tool that drafts, suggests and summarises alongside the person using it and leaves every decision and final action to that person. Also a product brand of Microsoft and GitHub.

Chatbot

Conversational assistant · virtual assistant

A program that people talk to in ordinary language, by text or voice, and that replies in a conversation. Older chatbots follow scripts; today's are built on large language models.

Autonomy Levels

Levels of autonomy

A graded scale that describes how much an AI system may do without a person, from only suggesting to acting alone and reporting afterwards. No single standard scale exists for AI agents.

Human-in-the-LoopHITL

A design pattern where an AI system pauses at defined points so a person can review, correct or approve before it proceeds.

Human-on-the-LoopHOTL

An oversight model in which the AI system acts on its own while a person monitors it and can step in or stop it.

08Security & safety21

Lethal Trifecta

Three capabilities combined in one AI agent: access to private data, exposure to untrusted content, and a way to send data out. With all three present, data can be stolen.

Prompt Injection

An attack in which text the model reads, such as a web page, an email or a document, contains instructions that the model then follows as if they came from its user.

Indirect Prompt Injection

Prompt injection delivered through content the AI fetches or is given, such as a web page, email, document or tool result, with no contact between the attacker and the system.

Jailbreak

A prompt or technique that gets a model to ignore its built-in safety rules and produce content it was trained to refuse.

Data Exfiltration

The unauthorised transfer of data out of a system; with AI, typically an agent tricked into sending private information to an attacker.

Guardrails

Checks placed around a model that inspect what goes in and what comes out, and block, rewrite or escalate anything that breaks the rules.

Red Teaming

Deliberately attacking your own AI system, the way an adversary would, to find its failures before someone else does.

Sandbox

An isolated environment in which an AI agent can run code and take actions without being able to reach anything outside it.

Least Privilege

The principle of giving an AI agent only the access and permissions its current task needs, and nothing more.

Excessive Agency

The condition in which an AI agent has more tools, broader permissions or more freedom to act without approval than its task requires, so that a mistake or a manipulation can cause real damage.

Confused Deputy

Confused deputy problem

A security flaw in which a program that holds legitimate privileges is tricked by a less privileged party into using them on that party's behalf; AI agents that act on untrusted content are a prime case.

Tool Poisoning

Tool poisoning attack

An attack in which a tool offered to an AI agent carries hidden instructions in its description or metadata; the model reads them as part of its context and may follow them, while the user sees only the tool's name.

Data Poisoning

An attack that plants manipulated examples in the data a model learns from or retrieves, so that the model later behaves as the attacker intends: making targeted errors, favouring an outcome or responding to a hidden trigger.

Adversarial Attack

Adversarial example

A deliberate attempt to make an AI model fail by giving it input crafted for that purpose, often a change too small for a person to notice that leads the model to a confident wrong answer.

Shadow AI

The use of AI tools inside an organisation without the knowledge or approval of IT, security or management.

Deepfake

Audio, video or an image, generated or altered with AI, that shows a real person saying or doing something they never said or did, convincingly enough to be taken as genuine.

Personally Identifiable InformationPII

Information that identifies a specific person, alone or combined with other data: a name, an ID number, an email address, a face. It is mainly a US term; European and Turkish law use the broader concept of personal data.

Alignment

AI alignment

The work of making an AI system pursue the goals and respect the values its developers and users intend, including in situations nobody spelled out in advance.

AI Safety

The field concerned with preventing AI systems from causing harm, whether through misuse by people, through failures of the system itself or through wider effects on society.

Sycophancy

A model's tendency to tell people what they want to hear: agreeing, flattering and backing down instead of giving an accurate answer.

Slop

AI slop

Low-quality content generated with AI in large quantities and published or passed on with little human checking: articles, images, videos, reports and code that look finished and say little.

09Evaluation & quality13

Evals

Evaluations

Systematic tests that measure how well an AI system performs the tasks it was built for, run on the same fixed set of cases every time the model, the prompt or the data changes.

Benchmark

A public, standardised test, with a fixed set of tasks and a fixed scoring method, used to compare AI models with one another.

LLM-as-a-Judge

LLM judge · model-graded evaluation

An evaluation method in which a language model grades the output of another AI system against written criteria, so that open-ended answers can be scored at a volume human reviewers cannot reach.

Golden Dataset

Golden set · gold-standard dataset

A curated set of test cases, each with an input and an answer approved by experts, kept as the fixed reference against which an AI system is evaluated.

Ground Truth

Reference answer · gold label

The verified correct answer for a given input, established by measurement, records or expert judgement, against which the output of an AI system is compared.

Precision and Recall

Two measures of a system that picks items out of a larger set: precision is the share of the items it picked that were right, and recall is the share of the right items that it picked.

Human Evaluation

Human eval

Assessment of AI outputs by people, such as subject experts, trained annotators or end users, who rate, compare or correct them according to defined criteria.

Observability

AI observability · LLM observability

The ability to see what an AI system in production actually did and why: each request, its steps, its cost, its speed and the quality of the result are recorded so that problems can be found and explained.

Tracing

Trace · LLM tracing

The step-by-step record of how an AI system handled one request or one agent run: every prompt, model call, retrieved document, tool call and output, with timings and cost.

Model Drift

Model decay

The loss of a model's quality in use over time, because the data it meets or the world it describes has moved away from what it was trained and tested on, or because the model behind the service has changed.

Bias

AI bias · algorithmic bias

A systematic skew in a model's results: errors that lean in one direction and, in the sense that matters most to organisations, outcomes that are consistently worse for some groups of people than for others.

ExplainabilityXAI

Explainable AI

The degree to which people can be given understandable reasons for a model's output: which inputs and factors led to this result, in terms that the person affected or the person responsible can act on.

Overfitting

A model that has learned its training data too closely, including the noise and coincidences in it, so that it performs very well on the examples it has seen and poorly on new ones.

10Infrastructure & economics15

GPU

Graphics processing unit

A processor built to carry out a very large number of simple calculations at the same time; designed for graphics, it became the standard hardware for training and running AI models.

Compute

Computing power

The processing power used to train and run AI models, supplied mostly by GPUs and measured in units such as GPU-hours; with data and algorithms, one of the three basic inputs of AI.

Inference Cost

The recurring cost of running a trained model: it is paid on every request, in tokens or in GPU time, and grows with the number of users, the length of prompts and the length of answers.

Token Pricing

Pay-per-token pricing

The standard way AI models are sold through an API: a price per million tokens, charged separately for the text sent in and the text generated, with output costing several times more.

Latency

Response time

The time between sending a request to a model and receiving its response; for a language model it is the wait for the first token plus the time needed to generate all the rest.

Time to First TokenTTFT

The time from sending a request until the first token of the model's response arrives; the main measure of how fast an AI application feels.

Rate Limit

A cap set by a model provider on how many requests and tokens a customer may send within a period, usually a minute; requests above it are rejected until capacity frees up.

Streaming

Response streaming

Delivering a model's response piece by piece as it is generated, so that the user sees the first words within moments and does not have to wait for the complete answer.

Batch Processing

Batch API · batch inference

Submitting many model requests as one job that is completed within hours, in exchange for a lower price and separate, higher volume limits.

Model Router

LLM router · model routing

A component that decides, for each request, which model should handle it: easy work goes to small, cheap models and the most capable model is kept for the requests that need it.

AI Gateway

LLM gateway

A single controlled entry point through which all of an organisation's applications reach AI models, where keys, limits, budgets, logging, guardrails and fallback are applied in one place.

LLMOps

LLM operations

The practices and tools for running applications built on large language models in production: versioning prompts and models, evaluating changes, tracing behaviour, controlling cost and releasing safely.

On-premise / Self-hosted

On-prem · self-hosting

Running AI models on infrastructure the organisation controls, in its own data centre or its own cloud account, so that prompts and data never pass through a model provider's shared service.

Edge AI

On-device AI

AI that runs on the device where the data arises, such as a phone, a laptop, a car, a camera or a machine on the factory floor, with no need to send each request to a data centre.

Open Source vs. Closed Model

Open vs. proprietary models

The choice between models whose weights you can download and run yourself and models offered only as a provider's service; it settles who controls the model, the data, the cost and the pace of change.

No term matches that search.