The AI Glossary

Accuracy is the proportion of correct predictions a model makes relative to the total number of predictions evaluated.

Agent Washing

Working knowledgeAgentic Protocols and Infrastructure

Agent washing is the practice of selling ordinary automation or a scripted chatbot as an autonomous AI agent.

Agentic AI

Agentic AI is the design approach in which software plans a sequence of steps, uses external software tools and takes autonomous action to execute multi-step goals, rather than generating a single response to a single prompt.

AI Agent

Working knowledgeAgentic Protocols and Infrastructure

An AI agent is a software system that uses an AI model to plan and execute multi-step tasks autonomously, with minimal human intervention.

AI Bubble

The AI bubble is the concern that AI-related investment, valuations and infrastructure spending have grown faster than the revenue and profits actually being generated, echoing the dot-com bubble of the early 2000s.

AI Due Diligence

AI due diligence is the work of establishing whether a company's AI claims, capabilities, costs and risks are what they are said to be, carried out before an acquisition or an investment.

AI Governance

AI governance is the set of frameworks, policies, controls and oversight mechanisms an organisation puts in place to manage risk, ensure compliance and maintain accountability across its use of artificial intelligence.

AI Readiness

AI readiness is whether an organisation's data, processes and people are in a state where AI can be deployed usefully, as distinct from whether the technology itself works.

AI Safety

AI safety is the interdisciplinary field focused on minimising the systemic risks, operational failures and societal harms caused by artificial intelligence systems.

AI Sandbox

An AI sandbox is a restricted, isolated environment used to test or run an AI system separately from live systems, so that unexpected or harmful behaviour can be observed and contained before it affects real operations, data or customers.

AI Slop

AI slop is the flood of low-quality, mass-produced content generated cheaply by AI and published with little or no human oversight, covering text, images and video.

AI Washing

AI washing is the deceptive practice of exaggerating, misrepresenting or falsely claiming that artificial intelligence is used within a company's products, services or internal operations.

AI Watermarking

AI watermarking is the practice of embedding subtle, mathematically detectable signals or metadata into text, images, audio or video generated by a model.

Algorithm

An algorithm is a precise, unambiguous sequence of mathematical and computational rules designed to process input data, perform calculations and solve a specific problem.

Algorithmic Bias

Algorithmic bias is the systematic production of prejudiced or unfair decisions by an artificial intelligence system, disadvantaging specific demographic groups, customers or operational segments.

API (Application Programming Interface)

An API, or application programming interface, is a defined interface that allows software systems to exchange information.

Artificial General Intelligence (AGI)

Artificial general intelligence (AGI) is a hypothetical AI capable of understanding or learning any intellectual task that a human being can, aiming to match the cognitive range of a person rather than perform one narrow task well.

Artificial Intelligence (AI)

Artificial intelligence is the branch of computer science dedicated to creating hardware and software systems capable of performing tasks that have historically required human intelligence.

B

3 terms

Benchmark

An AI benchmark is a standardised dataset, test suite or performance metric used to compare and rank artificial intelligence models on a specific capability, such as logical reasoning, language understanding or coding accuracy.

Benchmark Gaming

Benchmark gaming is the practice of deliberately optimising or fine-tuning models to excel on public evaluation datasets without achieving any generalisable improvement in the underlying capability.

Black Box (AI)

Working knowledgeTechniques and Architectures

A black box, applied to AI, is a system whose internal decision-making cannot be inspected or explained, even by the people who built it.

C

7 terms

Chain-of-Thought

Chain-of-thought is a model working through intermediate steps before it gives an answer, rather than producing the answer in a single move.

Chatbot

FoundationsAI Economics and Infrastructure

A chatbot is a software application designed to simulate human conversation through text or voice.

Compute

Compute is the raw hardware processing power, measured in floating-point operations per second (FLOPS), required to train, fine-tune and run artificial intelligence models.

Computer Vision

Computer vision is the field of artificial intelligence that enables software to acquire, process, analyse and comprehend visual information from digital images, video streams and sensor data.

Context Engineering

Context engineering is the discipline of assembling the right information, tools and instructions into a model's context window before it acts.

Context Window

The context window is the total volume of data a model can process at one time.

Copilot

FoundationsAI Economics and Infrastructure

A copilot is an AI assistant built into the tool someone already works in, drafting and suggesting while the person using it keeps control of what is accepted.

D

8 terms

Data Centre

A data centre is a dedicated physical facility housing centralised computing servers, networking equipment and enterprise storage.

Data Governance

Data governance is the management framework covering the availability, usability, integrity, privacy and security of an organisation's data assets.

Data Privacy

Data privacy in artificial intelligence is the set of practices and legal obligations governing the collection, processing, storage and sharing of personally identifiable information.

Data Residency

Data residency is the legal and regulatory requirement that personal or corporate data be collected, processed and stored within specific geographic or legal jurisdictions.

Deep Learning

Deep learning is a subset of machine learning that uses neural networks with many layers.

Deepfake

A deepfake is synthetic media, including video, audio and images, manipulated or generated using deep neural networks to convincingly impersonate a real person.

Digital Twin

A digital twin is a virtual representation of a physical asset, facility, product or operational system, kept current by real-time sensor streams and AI predictive models.

Distillation

Distillation is a training technique where a smaller model learns to imitate a larger one, typically by training on the larger model's outputs rather than on raw data.

E

4 terms

Embeddings

Working knowledgeEvergreen Landmarks

An embedding is a numerical representation of meaning, produced by converting a piece of text or an image into a list of numbers positioned so that similar meanings sit close together.

EU AI Act

The EU AI Act, enacted in 2024, is the world's first comprehensive legally binding regulatory framework governing the development and commercial operation of artificial intelligence within the European Union.

Evals

Evals are systematic, repeatable tests that measure the quality of an AI system's output against a defined set of cases with known good answers.

Explainable AI (XAI)

Working knowledgeTechniques and Architectures

Explainable AI (XAI) is the set of methods and engineering techniques that make the internal decision-making pathways and outputs of complex machine learning models understandable to human domain experts.

F

3 terms

Fine-Tuning

Fine-tuning is the process of taking a pre-trained foundation model and updating its parameters by training it on a smaller, domain-specific dataset.

Foundation Model

A foundation model is a large machine learning model trained on vast, multi-modal datasets using self-supervised learning.

Frontier Model

A frontier model is a state-of-the-art foundation model that approaches or surpasses the existing technical limits of reasoning, multi-step problem solving and multimodal comprehension.

G

5 terms

Generative AI (GenAI)

Generative AI is the class of models, large language models among them, trained to create new content such as text, code or images by predicting sequence probabilities.

Generative Engine Optimisation (GEO)

Generative engine optimisation is the practice of making content likely to be retrieved, cited and recommended by AI answer engines such as ChatGPT, Perplexity and AI Overviews.

GPU

Working knowledgeAgentic Protocols and Infrastructure

A GPU (Graphics Processing Unit) is a specialised computer chip designed for rendering graphics that turned out to be extremely efficient at the parallel calculations AI models need, both for training them and for running them.

Grounding

Grounding is the practice of tying a model's answers to specific source material that can be checked, rather than to whatever it absorbed during training.

Guardrails

Guardrails are the rules and filters built around an AI system that constrain what it can produce, access or do.

H

4 terms

Hallucination

A hallucination is an AI output that states false information with complete confidence: invented citations, fabricated statistics, fictional case law.

Horizontal AI

Horizontal AI is a general-purpose model or platform designed to work across industries and use cases, rather than being built around the specifics of any one of them.

Human-in-the-Loop (HITL)

Human-in-the-loop is a workflow design in which a person reviews, approves or corrects AI output before it takes effect.

Human-on-the-Loop (HOTL)

Human-on-the-loop is a supervision model in which a person monitors an AI system's overall behaviour and intervenes only when a result crosses a defined threshold or is flagged as an exception, rather than reviewing every individual action.

I

1 term

Inference Cost

Inference cost is the cost of actually running an AI model to produce an output.

J

2 terms

Jagged Frontier

FrontierCore AI Concepts

The jagged frontier is the uneven capability boundary of modern foundation models, where a model executes extraordinarily complex analytical tasks with elite expertise while unexpectedly failing at simpler, adjacent operational ones.

Jailbreaking

Jailbreaking is the practice of using crafted prompts to bypass an AI model's built-in safety filters, ethical guardrails and operational alignment restrictions.

L

1 term

Large Language Model (LLM)

A large language model (LLM) is a foundation model with billions of parameters, trained on vast textual datasets using transformer neural network architectures.

M

7 terms

Machine Learning (ML)

Machine learning is the paradigm of training systems to identify patterns from data rather than programming them with explicit, hard-coded rules.

Model Collapse

FrontierAgentic Protocols and Infrastructure

Model collapse is the gradual degradation in quality that happens when AI models are repeatedly trained on data generated by earlier AI models, rather than on original human-made data.

Model Context Protocol (MCP)

The Model Context Protocol (MCP) is an open standard that allows software systems, including AI agents, to connect to external tools and data sources through a single, consistent interface.

Model Drift

Model drift is the gradual decay in a model's performance as the world it operates in moves away from the data it was trained on.

Model Risk

FrontierAgentic Protocols and Infrastructure

Model risk is the risk that decisions taken on incorrect outputs from a model cause financial, regulatory or reputational harm.

Multi-Agent Orchestration

Multi-agent orchestration is an architecture in which multiple AI agents, often a mix of models with different capability and cost, are coordinated to complete a task together rather than relying on one model for everything.

Multimodal AI

Multimodal AI is a model built from the start to take in, cross-reference and generate several distinct data types at the same time, including text, images, audio, video and numerical tables.

N

3 terms

Narrow AI

Narrow AI is a system designed to perform a specific, bounded task, such as forecasting churn or parsing contracts.

Natural Language Processing (NLP)

Natural language processing (NLP) is the field concerned with enabling computers to comprehend, interpret, evaluate and generate human languages.

Neural Network

A neural network is a computing structure made of interconnected nodes organised in layers, which process data using mathematical functions.

O

4 terms

Observability

Observability is the ability to see, trace and explain what an AI system actually did.

Open-Source AI

Open-source AI refers to models, libraries and frameworks whose underlying source code, architecture and training scripts are made publicly available for unrestricted commercial use, modification and distribution.

Open-Weight Models

Open-weight models are models whose parameters are published for anyone to download, inspect, fine-tune or run on their own infrastructure.

Optical Character Recognition (OCR)

Optical character recognition (OCR) is computer vision technology that converts printed, handwritten or scanned text inside digital images and PDF files into machine-encoded, searchable text.

P

3 terms

Predictive AI

FoundationsTechniques and Architectures

Predictive AI is a class of model that analyses historical data to forecast future outcomes.

Prompt Engineering

Prompt engineering is the practice of deliberately structuring the input given to an AI model.

Prompt Injection

Prompt injection is a security vulnerability where an attacker manipulates a language model's input instructions to override system rules, safety filters and operational boundaries.

R

7 terms

Reasoning Model

A reasoning model is a class of AI model that works through a problem step by step before producing an answer, rather than generating a response immediately.

Recommendation System

A recommendation system is software that predicts user preferences and suggests relevant items, content or services.

Red Teaming

Working knowledgeTechniques and Architectures

Red teaming is adversarial security testing in which specialists deliberately attempt to breach an AI system's guardrails, bypass its alignment restrictions and trigger harmful outputs.

Reinforcement Learning (RL)

Reinforcement learning is a way of training a system by trial and error, rewarding the actions that work and penalising the ones that do not, until it learns a sequence of decisions that scores well.

Responsible AI

FrontierAgentic Protocols and Infrastructure

Responsible AI is the governance framework that ensures AI systems are developed, deployed and monitored ethically, safely and transparently.

Retrieval-Augmented Generation (RAG)

Retrieval-augmented generation is an architectural pattern that connects a generative model to a specific body of content, typically proprietary company documents, rather than relying on the model's public training alone.

Robotic Process Automation (RPA)

Robotic process automation is a set of software tools that configure rule-based bots to carry out repetitive, structured digital administrative tasks.

S

9 terms

Scheming

Scheming is when an AI system behaves as though aligned with its developers' goals while it is being tested, but would pursue a different goal once confident it is not being monitored.

Sentiment Analysis

Sentiment analysis is a natural language processing technique that reads unstructured text to identify and quantify emotional tone, customer attitude and subjective opinion.

Shadow AI

Shadow AI is the use of AI tools inside an organisation without the knowledge or approval of those responsible for security and risk.

Small Language Model (SLM)

A small language model is a compact model chosen in place of a frontier one because it costs less to run, answers faster and is small enough to host on hardware you control.

Sovereign AI

Sovereign AI is the principle that a nation or organisation keeps control over where AI runs and where the underlying data lives, rather than ceding it to foreign providers or systems it cannot inspect or govern.

Speech Recognition

Speech recognition is the technology that converts spoken audio into machine-readable text.

Superintelligence

Superintelligence is a hypothetical AI that would vastly exceed human ability across every field at once, not merely match it.

Synthetic Data

Synthetic data is artificially generated data produced by algorithms and generative models rather than collected from real-world events.

System Prompt

A system prompt is the standing set of instructions a deployer gives a model, sent ahead of every conversation and invisible to the people using it.

T

8 terms

Test-Time Compute

Test-time compute is the practice of allocating extra computation at the moment a model answers rather than during training, letting it explore several reasoning paths or search trees before committing to a response.

Text-to-Image

Text-to-image is a class of generative AI system that turns a written description into a picture.

Token

Working knowledgeAgentic Protocols and Infrastructure

A token is the unit by which AI models read and generate text.

Tool Use

Tool use is a model calling external software during a task: querying a database, sending an email, running a calculation, or reaching another system through its API.

Total Cost of Ownership (TCO)

Total cost of ownership for AI is the full cost of acquiring, deploying and running a system across its operational life.

Training Data

Working knowledgeEvergreen Landmarks

Training data is the dataset a machine learning model learns from.

Transformer

The transformer is a neural network architecture that relies entirely on self-attention to process sequential data in parallel rather than one element at a time.

Turing Test

FoundationsEvergreen Landmarks

The Turing test is a historical benchmark asking whether a machine can exhibit human-equivalent conversational intelligence.

V

4 terms

W

1 term

Workflow Automation