The A-Z of Artificial Intelligence

Artificial Intelligence (AI)

A branch of computer science focused on building systems that can perform tasks which typically require human intelligence, such as reasoning, recognising patterns, learning from experience, and making decisions. Modern AI ranges from narrow tools designed for a single task to large models capable of generating text, images, code, and more.

Artificial General Intelligence (AGI)

A form of AI capable of understanding, learning, and applying knowledge across a broad range of tasks at a level comparable to human cognition. Whether current frontier models represent early forms of AGI is actively debated, with major AI labs holding divergent views on how to define it and how close we are. Some commercial definitions, such as the one in OpenAI's partnership with Microsoft, are now tied to revenue thresholds rather than capability tests, which gives the term a fluid quality outside research circles.

Artificial Neural Network (ANN)

A machine learning model loosely inspired by the structure of the human brain, forming the foundation of many deep learning systems.

Algorithm

A set of step-by-step rules or instructions used by AI systems to process information and solve problems.

Agents

AI systems that can act independently to complete tasks, make decisions, and interact with tools or services on behalf of users. Rather than responding to a single prompt, agents plan across multiple steps, use external tools such as web search or code execution, and adapt their approach based on results. Orchestrating multiple agents to work together on complex tasks is an increasingly common pattern in production AI systems.

AI Browser

A web browser that uses AI to read, summarise, navigate, or act on pages on the user's behalf. Examples include OpenAI's Atlas and Perplexity's Comet, with similar agentic features appearing in other browsers through extensions and integrations. Rather than the user driving every click, an AI browser can plan a task, visit several sites, compare information, and surface or execute a recommendation. The shift challenges how brands appear to a customer, since the agent now sits between them.

Read more: When the browser becomes the destination

AI Pilot

A time-bounded test of an AI application in a real working environment, designed to answer a specific question before any wider commitment. A good pilot is scoped around a business problem, not a technology demonstration. The measure is whether it changes how people work, not whether the technology runs. Most enterprise AI pilots fail to reach production because they are designed to prove capability rather than to establish operational value.

Read more: What a Good AI Pilot Actually Looks Like

Agentic AI

AI systems designed to act autonomously over time, rather than respond to a single prompt and stop. An agentic system can plan across multiple steps, use external tools, call APIs, and adapt its approach based on what it finds, operating more like a process than a conversation. The shift from single-turn AI to agentic systems changes the questions organisations need to ask about oversight, reliability, and accountability in production.

AI Governance

The policies, processes, and oversight structures an organisation puts in place to manage how AI is used. Governance covers who owns AI decisions, how models are audited for accuracy and bias, what happens when an AI system causes harm, and how the organisation meets regulatory requirements such as the EU AI Act. As AI moves from experiment to operational infrastructure, governance is shifting from a compliance checkbox to a board-level concern.

Alignment

The challenge of ensuring AI systems pursue goals that are aligned with human values and intentions. It is an active and unsolved area of research, and is one of the central reasons safety teams exist inside frontier AI labs.

Automation

Using technology to perform tasks with minimal human input, often accelerated by AI.

Backpropagation

A training method for neural networks that adjusts weights by calculating gradients of error.

Benchmarking

The practice of testing AI systems against standardised tasks such as MMLU, BIG-bench, or ARC. Benchmarks help compare model capabilities, track improvements, and evaluate performance across different use cases.

Bias (AI Bias)

Systematic distortion in AI outputs caused by biased training data or flawed assumptions. Bias has become a regulatory and reputational concern, particularly in hiring, lending, and other high-impact decisions, and is one of the areas the EU AI Act specifically addresses.

Big Data

Massive datasets used by AI to identify patterns, often beyond human processing capacity. The term peaked in the 2010s and is heard less often now that the same capability is assumed in most modern AI systems.

ChatGPT

A conversational AI product developed by OpenAI, built on large language models. It can engage in dialogue, answer questions, write and edit text, analyse data, and assist with a wide range of tasks through a chat interface.

Claude

An AI assistant developed by Anthropic, available as a family of models ranging from fast and lightweight to its most capable versions. Claude is designed with a focus on safe and predictable behaviour, and supports extended reasoning, tool use, and long-context tasks. It is one of the leading frontier models alongside models from OpenAI and Google.

Computer Use

A capability that lets an AI model control a computer through a screen, mouse, and keyboard rather than through a dedicated API. Anthropic shipped the first commercial version of this in October 2024, with OpenAI's Operator following in January 2025. It extends agentic AI beyond web browsing into desktop applications and legacy systems that have no public API, with the trade-off of being slower and more error-prone than direct integration.

Computer Vision

AI systems that analyse and interpret visual inputs such as images or videos.

Convolutional Neural Network (CNN)

A neural network particularly effective for processing visual data.

Compute

The processing power required to train and run AI models, typically measured in FLOPs (floating-point operations per second). Compute is the primary cost driver in frontier model development, and the availability of it through chips, data centres, and cloud infrastructure is one of the main constraints on how fast the field can move. Control over compute supply chains has become a significant geopolitical concern, with export restrictions on advanced chips increasingly shaping which organisations and countries can build frontier models.

Context Engineering

The practice of structuring the information provided to an AI model, including system instructions, conversation history, retrieved documents, and tool outputs, to produce reliable and relevant results. As AI systems become more complex, how a model is given context has become as significant as the underlying model itself. Sometimes used as a more precise alternative to prompt engineering.

Context Window

The amount of input data an AI model can process in a single exchange. Long-context models can handle entire books, codebases, or multi-hour transcripts. The size of the context window affects how much information a model can draw on when generating a response.

Conversational AI

AI designed to engage in dialogue, including chatbots and virtual assistants.

Data Mining

The process of discovering patterns and extracting useful information from large datasets using AI methods.

Deep Learning

A subset of machine learning using layered neural networks to model complex patterns in data.

DeepSeek

A family of AI models developed by a Chinese research organisation that attracted significant attention in early 2025 when its R1 reasoning model demonstrated performance comparable to leading US frontier models at a fraction of the reported training cost. DeepSeek has continued to release updated models in both its V3 and R1 lines, and remains one of the most-watched open-weights labs alongside Meta and Mistral. The original release reshaped assumptions about the relationship between compute investment and model capability.

Diffusion Models

Generative models that iteratively transform noise into coherent outputs, widely used in image generation.

Domain Adaptation

A technique allowing AI models trained in one domain to perform well in a different, but related, domain.

Embeddings / Vector Database

An embedding is a numerical representation of a piece of text, image, or other content that captures its meaning in a way machines can compare. A vector database stores these embeddings so an AI system can quickly find the most relevant pieces of content for a given query. Together, they are the underlying mechanism behind retrieval-augmented generation, semantic search, and personalised recommendations in modern AI products.

Embodied AI

AI systems that interact with the physical world through sensors and actuators, often used in robotics.

Ethics (AI Ethics)

Principles that guide the responsible design, deployment, and use of AI, including fairness, safety, and transparency.

EU AI Act

Legislation introduced by the European Union that regulates AI systems according to their potential risk. High-risk applications, such as those used in hiring, credit scoring, or critical infrastructure, face requirements around transparency, human oversight, and data governance. It entered into force in August 2024. Prohibitions on the most harmful uses of AI applied from February 2025, general-purpose AI obligations from August 2025, and obligations on high-risk systems from August 2026. It is the most significant piece of AI-specific regulation in force globally.

Expert System

An AI that mimics the decision-making ability of a human expert by applying predefined rules to a knowledge base.

Explainability

The ability to understand and interpret how an AI system arrived at its conclusions or decisions.

Fine-Tuning

A method of customising a pre-trained AI model on a smaller, task-specific dataset.

Foundation Model

A large AI model trained on broad data and designed to be adapted for a wide range of downstream tasks. Foundation models serve as the base for many commercial AI products and services, fine-tuned or prompted to meet specific use cases rather than trained from scratch for each application.

Frontier Model

The current generation of large AI models at the leading edge of capability, typically produced by a small number of labs with significant compute and data resources. The term is informal and the frontier moves frequently. Companies including OpenAI, Anthropic, Google, and xAI currently sit in this category, with DeepSeek operating at similar capability in the open-weights tier. Frontier model is often used in regulatory and safety conversations to distinguish the most capable systems from smaller, more constrained ones.

Frugal AI

The practice of designing efficient, cost-effective, and environmentally responsible AI systems.

Function Calling

A capability that allows AI models to trigger external tools, APIs, or predefined code functions as part of generating a response. This is widely used in agentic and autonomous systems to extend a model's abilities beyond text generation.

Fuzzy Logic

A form of logic that handles imprecision, enabling AI to make decisions based on approximate information.

Generative AI (GenAI)

AI that produces new content, including text, images, audio, and video, based on learned data patterns. Generative AI is the category that drove the post-2022 wave of commercial AI adoption, and is the form of AI most people encounter through products such as ChatGPT, Claude, and Gemini.

Generative Adversarial Network (GAN)

A system of two neural networks competing to generate increasingly realistic data.

Generative Pre-trained Transformer (GPT)

A language model architecture trained on large text corpora to predict and generate coherent language.

Grounding

The practice of anchoring a model's responses to specific, verifiable information rather than relying solely on what it learned during training. A common approach is retrieval-augmented generation, where relevant documents or data are provided alongside the prompt so the model can draw on them directly.

Guardrails

Predefined constraints that keep AI systems within safe or acceptable limits. Guardrails help prevent harmful or inappropriate responses, enforce compliance, and align model behaviour with user expectations or legal requirements.

Hallucination

When an AI generates content that is plausible-sounding but factually incorrect or unsupported by its training data or the context provided. Hallucination cannot be fully eliminated, but it can be reduced through grounding, retrieval-augmented generation, and careful evaluation of outputs in production.

Human-in-the-Loop

An approach where humans provide oversight, validation, or feedback to AI systems during or after decision-making.

Read more: The Trouble with Human in the Loop

Hyperparameter

A setting used to control the training process of an AI model, such as learning rate or layer size.

Inference

Using a trained AI model to generate outputs or predictions from new input data. Training is a one-time cost and inference is the ongoing one, which is why the economics of running AI at scale are often dominated by inference rather than the headline training figures that get the attention.

Intelligent Agent

An autonomous AI that perceives its environment and takes actions to achieve specific goals.

Internet of Things (IoT)

A network of connected devices that collect and exchange data, often enhanced by AI for smarter operations.

Joint Embedding Predictive Architecture (JEPA)

A model that predicts relationships between different data embeddings, especially effective for multimodal or temporal tasks.

Job Displacement

A potential social impact of AI where automation replaces human roles in certain industries.

Knowledge Graph

A structured way to represent facts and their relationships, used by AI to reason and understand context.

Knowledge Distillation

A technique where a smaller model learns to replicate the behaviour of a larger, more complex one.

Large Language Model (LLM)

An AI model trained on large amounts of text data, capable of generating coherent and contextually appropriate language. LLMs are the foundation of most current generative AI products, including ChatGPT, Claude, and Gemini, and are the model type meant when people talk about AI without further qualification.

Low-Rank Adaptation (LoRA)

A fine-tuning technique that updates only a small number of parameters, improving efficiency.

Logic Programming

A programming method where logic-based rules determine actions or decisions.

Machine Learning (ML)

A subset of AI where systems learn patterns from data rather than being explicitly programmed.

Mixture of Experts (MoE)

A model architecture that contains several specialised sub-networks, with a routing mechanism that decides which sub-networks to use for each input. Only a fraction of the model is active for any given request, which makes large MoE models faster and cheaper to run at inference than their total parameter count suggests. Notable models using this approach include DeepSeek V3 and Mistral's Mixtral family. The architecture is now common in large language models, though developers do not always disclose whether their model uses it.

Model Context Protocol (MCP)

An open standard developed by Anthropic that defines how AI models connect to external tools, data sources, and services. MCP allows developers to build integrations once and have them work across different AI systems. It has been widely adopted as a practical way to extend what AI models can do without building bespoke connectors for each application.

Read more: Model Context Protocol (MCP): A Framework for Connecting AI and Business Systems

Model Training

The process of teaching an AI model to recognise patterns by exposing it to data and feedback.

Multimodal AI

AI systems that can process and relate different kinds of data, such as text, images, and sound.

Natural Language Processing (NLP)

The field of AI focused on enabling machines to understand, interpret, and generate human language.

Natural Language Understanding (NLU)

A subfield of NLP concerned with extracting meaning, intent, and context from text or speech.

Neural Network

A series of algorithms that recognise relationships in data, loosely inspired by the structure of the human brain.

OpenAI

An AI company founded in 2015 and now one of the largest AI companies in the world. OpenAI operates the GPT model family and ChatGPT, and has a deep commercial and infrastructure partnership with Microsoft. It is one of the leading frontier model developers alongside Anthropic, Google, and xAI.

Orchestration

The coordination of multiple AI models, agents, or tools working together to complete a task. An orchestration layer manages the flow of information between components, decides which model or tool to use at each step, and handles errors or unexpected outputs. As agentic AI systems become more common in production, orchestration has become a significant area of both engineering and product design.

Open Weights

An AI model whose trained parameters are released publicly, allowing anyone to download, run, fine-tune, or adapt the model without going through the original developer. Notable open-weights models include Meta's Llama family, DeepSeek's V3 and R1, Mistral's models, and Alibaba's Qwen. Open weights is distinct from open source: most open-weights releases publish the weights but not the training code, data, or process. The strategic question for a buyer is usually whether the freedom to self-host justifies the work of running the infrastructure.

Overfitting

When a model performs well on training data but poorly on unseen data due to excessive specialisation.

Optimisation

Improving AI model performance by adjusting parameters, architecture, or training data.

Pattern Recognition

The ability of AI systems to detect regularities or structures in data.

Planning (AI Planning)

The ability of AI agents to break down high-level instructions into step-by-step tasks or sequences. Planning allows models to operate over time, manage dependencies, and pursue goals with limited human oversight.

Predictive Analytics

AI-powered analysis of data to forecast trends or future events.

Prompt Engineering

Crafting input prompts to guide or control the outputs of a generative AI system.

Pruning

Reducing model complexity by removing redundant components or connections.

Quantisation

Reducing the precision of model parameters to improve computational efficiency, often with acceptable trade-offs in performance.

Query

A structured or natural language request to an AI system to retrieve or generate information.

Reinforcement Learning

A training method where an AI learns by trial and error, receiving rewards or penalties based on its actions.

RLHF (Reinforcement Learning from Human Feedback)

A training technique that uses human preferences to fine-tune AI models. Human evaluators rank model outputs, and the model learns to favour responses that align with those judgements. RLHF is a key method used to make large language models more helpful, harmless, and aligned with user intent.

Reasoning

The process by which AI systems derive conclusions, explanations, or decisions from available information. See also: Reasoning Models.

Reasoning Models

A category of AI model that generates extended internal reasoning before producing a final response. Rather than answering immediately, these models work through a problem step by step, which improves performance on tasks that benefit from structured thinking such as mathematics, coding, and complex analysis. Once a distinct category led by OpenAI's o-series and DeepSeek's R1, reasoning capability is now built into most frontier models, including Claude, Gemini, and Grok. The underlying technique is sometimes called test-time compute scaling.

Retrieval-Augmented Generation (RAG)

Combining generative models with external information sources to help improve factual accuracy and reduce hallucination. RAG has become one of the most common patterns in production AI systems, used to ground a model in a specific knowledge base such as a company's internal documents, product catalogue, or policy library.

Robotic Process Automation (RPA)

Automating routine business tasks using software bots, often enhanced by AI.

Supervised Learning

A method where AI is trained on labelled datasets to learn specific outputs for given inputs.

Small Language Model (SLM)

A compact AI model designed to run efficiently on devices with limited compute, such as laptops, phones, or edge hardware. While smaller than frontier models, recent SLMs have demonstrated strong performance on focused tasks. Their lower cost and ability to run locally make them practical for use cases where sending data to a cloud-based model is not feasible or appropriate.

Shadow AI

The use of AI tools by employees without the knowledge or approval of IT, security, or leadership. Shadow AI is the AI equivalent of shadow IT. It emerges when sanctioned tools do not meet people's needs and they reach for alternatives. The risks include data privacy exposure, intellectual property leakage, and compliance gaps, particularly where sensitive information is being pasted into consumer AI products. Most organisations discover the scale of it only after putting a policy in place.

Steerability

The ability to guide or constrain an AI model's behaviour according to user intent or parameters.

Stable Diffusion

A widely used open-source image generation model that creates images from text prompts using diffusion-based techniques.

Superintelligence

A hypothetical future AI that surpasses human intelligence in all domains.

Synthetic Data

Artificially generated data used to train, fine-tune, or evaluate AI models. It can be used to simulate edge cases, balance biased datasets, or avoid reliance on private or sensitive real-world data.

Transfer Learning

Using a model trained on one task as a starting point for another related task.

Transformer

A neural network architecture that enables models to understand relationships in sequential data. Forms the basis for LLMs like GPT.

Test-Time Compute

The processing power used by an AI model at the point of generating a response, as distinct from the compute used during training. Reasoning models increase test-time compute deliberately, spending more processing time working through a problem before responding. This represents a shift in how AI capability is developed: rather than always training larger models, additional thinking at inference time can improve results on specific task types.

Tokens / Tokenisation

The unit of text that an AI model actually processes. A token is roughly three to four characters of English text, so a 1,000-word document is around 1,300 tokens. Model pricing, context window limits, and processing speed are all measured in tokens rather than words or characters. Understanding this matters because token counts differ significantly between languages and content types, and they directly affect the cost of running an AI system at scale.

Turing Test

A measure of a machine's ability to exhibit intelligent behaviour indistinguishable from a human. Proposed by Alan Turing in 1950 and considered a foundational thought experiment, it is no longer treated as a meaningful benchmark, since modern language models routinely pass informal versions of it without being broadly considered intelligent in the sense Turing intended.

Unsupervised Learning

Training AI on data without labelled outputs, allowing it to find patterns or structures independently.

User Experience (UX) AI

The use of AI to personalise, adapt, or improve digital interfaces and user interactions.

Video Generation

AI models that produce moving images from a text prompt, image, or short input clip. Examples include OpenAI's Sora, Google's Veo, Runway, and Kling. Video generation is more compute-intensive and slower than image generation, with output lengths typically measured in seconds rather than minutes. Quality has advanced quickly through 2025 and 2026, raising fresh questions about content provenance, consent, and the cost of producing video for commercial purposes.

Virtual Assistant

An AI-powered tool that performs tasks or provides information through voice or text, such as Alexa or Siri.

Vision Transformer (ViT)

A model architecture that applies transformer principles to computer vision tasks with strong performance.

Weak AI

Also called narrow AI. Systems designed for specific tasks, such as image classification or recommendation.

Weights

Numerical parameters in neural networks that are updated during training to influence learning outcomes.

eXplainable AI (XAI)

AI systems designed to make their decisions and logic understandable to humans.

YOLO (You Only Look Once)

A real-time object detection algorithm that identifies multiple objects in images or video in a single pass.

Zero-shot Learning

The ability of AI models to perform tasks they have not been explicitly trained for, by generalising from related knowledge.