Free resource

Every word people use about AI, explained.

245 terms in plain English, including the ones people often get wrong. No sign-up needed.

AI regulation

AI Security Institute: The UK government's research organisation on serious risks from advanced AI, part of DSIT. Renamed from the AI Safety Institute in February 2025. It tests and studies frontier AI; it is not a regulator.
AI literacy: The skills, knowledge and understanding to use AI knowingly and to be aware of its risks. The EU AI Act requires providers and deployers to take measures to support it among staff; since the 2026 Omnibus, no specific level has to be guaranteed.
Automated decision-making: A decision about a person taken by technology with no meaningful human involvement. In the UK, Articles 22A to 22D of UK GDPR (in force 5 February 2026) set the safeguards for significant decisions of this kind.
Conformity assessment: The process of showing that a high-risk AI system meets the EU AI Act's requirements before it is placed on the market. Depending on the system, it is done by the provider itself or involves an independent notified body.
DPIA: Data protection impact assessment. A UK GDPR requirement before processing likely to result in high risk to people. The ICO says most AI use involving personal data will need one.
Deployer (AI Act): An organisation using an AI system under its own authority, other than for personal non-professional use. Most businesses buying AI are deployers, with their own duties for high-risk systems.
Digital Omnibus on AI: Regulation (EU) 2026/1744, in force from 27 July 2026. It amended the AI Act: high-risk obligations delayed to December 2027 and August 2028, the AI literacy duty softened, SME relief extended, and a new prohibition added.
GPAI model with systemic risk: A general-purpose AI model with high-impact capabilities, presumed where training compute exceeds 10^25 floating point operations. Its provider must assess and mitigate systemic risks on top of the standard GPAI duties.
General-purpose AI model: An AI model trained on broad data that can perform a wide range of distinct tasks and be built into many downstream systems. Large language models are the obvious example. Its provider has specific EU AI Act duties.
High-risk AI system: Under the EU AI Act, an AI system used in a sensitive area listed in Annex III (such as hiring, credit scoring or education) or as a safety component of a product covered by Annex I legislation. Carries the Act's heaviest obligations.
ISO/IEC 42001: The international standard for an AI management system, published December 2023. Organisations can be certified against it; ISO/IEC 42006 sets the rules for the bodies that certify.
NIST AI RMF: The US National Institute of Standards and Technology's AI Risk Management Framework, released January 2023. Voluntary, and organised into four functions: Govern, Map, Measure and Manage.
Principles-based regulation: Regulating by outcomes rather than detailed rules. The UK's approach to AI: five non-statutory principles applied by existing regulators in their own sectors, instead of a single AI law.
Prohibited AI practices: Uses the EU AI Act bans outright under Article 5, such as social scoring, harmful manipulation, untargeted facial image scraping and emotion recognition at work or in education. Banned since 2 February 2025.
Provider (AI Act): The person or organisation that develops an AI system or GPAI model, or has one developed, and places it on the market or puts it into service under its own name or trademark. Carries most AI Act obligations.
Regulatory sandbox: A supervised setting where firms can test products under a regulator's oversight, sometimes with specific rules relaxed for a limited time. Examples include the FCA's AI Live Testing, the MHRA AI Airlock and the proposed AI Growth Lab.

Basics

AGI (artificial general intelligence): A hypothetical AI that could match or exceed people across most intellectual tasks. There is no agreed definition or test, and forecasts of when, or whether, it will arrive vary widely. Treat confident claims in either direction with care.
AI (artificial intelligence): Computer systems that do tasks we would normally say need human intelligence, such as understanding language, recognising images or making predictions. The term dates from a 1955 proposal for a research workshop at Dartmouth College. Today it covers everything from spam filters to chatbots, so ask which kind of AI someone means.
AI assistant: A general-purpose chat tool built on a large language model that can write, answer questions, analyse files and, increasingly, take actions such as searching the web. The same assistant may behave differently depending on its plan, settings and which model is selected.
Algorithm: A set of step-by-step instructions for solving a problem. In machine learning, the training algorithm is the recipe for learning from data, and the model is what that recipe produces. People often say algorithm when they mean model.
Chatbot: A program you talk to in a conversation. Older chatbots followed scripted decision trees; modern ones use large language models, so they can handle open questions but can also give confident wrong answers. Treat what they say as a draft to check.
Computer vision: The field of AI that interprets images and video: reading text in photos, spotting defects on a production line, counting cars or describing a scene. Accuracy depends heavily on whether the images look like the ones the system was trained on.
Context window: The maximum amount of text, measured in tokens, a model can take into account at once, including your prompt, any files, the conversation so far and its own reply. Beyond it, earlier material is dropped or summarised. A big window does not guarantee the model uses everything in it equally well.
Deep learning: Machine learning using neural networks with many layers. The layers let the model build up from simple patterns to complex ones, which is why deep learning drives modern speech, image and language systems. It needs a lot of data and computing power.
Generative AI: AI that creates new content, such as text, images, audio, video or code, in response to a request. It works by producing output that is statistically likely given its training, which is why it can be fluent and wrong at the same time.
Hallucination: When an AI tool states something false as if it were true, such as an invented statistic, quote, case law or reference. It happens because the model generates plausible text rather than looking facts up. Some prefer the term confabulation. Check anything that matters against a source.
Knowledge cutoff: The date after which a model's training data stops, so it knows nothing about later events unless it can search the web or you give it the information. Models are not always reliable about their own cutoff, so check when recency matters.
LLM (large language model): A model trained on very large amounts of text to predict the next piece of text. That simple objective, at scale, produces systems that can answer questions, summarise, translate, draft and write code. LLMs power ChatGPT-style assistants.
Machine learning: The branch of AI where a system learns patterns from examples rather than following rules a person wrote by hand. Show it thousands of labelled emails and it learns to spot spam. Almost all modern AI, including chatbots, is built this way.
Model: The trained system at the heart of an AI tool: a very large set of learned numbers that turns an input into an output. The model is not the same as the product. One app can switch between several models, and one model can sit behind many apps.
NLP (natural language processing): The field of AI concerned with understanding and producing human language, from spell-checkers and translation to sentiment analysis and chatbots. Large language models are now the dominant approach within it.
Narrow AI: AI built for a specific task or range of tasks, such as recognising faces or recommending films. Every AI system in use today is narrow in this sense, even general-purpose assistants, which remain unreliable outside the patterns in their training.
Neural network: A model made of layers of simple connected units, loosely inspired by neurons in the brain. Each connection has a numerical weight, and training adjusts those weights until the network's outputs match the examples. The resemblance to a real brain is loose; it is maths, not biology.
Predictive AI: AI that forecasts or classifies rather than creates: predicting demand, flagging fraud, scoring credit risk or sorting support tickets. It has been in business use for far longer than generative AI and is often the better fit for structured data problems.
Prompt: What you type or say to an AI tool to get a result: a question, an instruction, or instructions plus material to work on. The clearer the task, context and expected format, the more useful the output tends to be.
Speech-to-text: Converting spoken audio into written text, also called automatic speech recognition. It powers meeting transcripts and voice dictation. Accuracy drops with background noise, crosstalk, strong accents the system has seen less of, and specialist vocabulary.
Text-to-image: Generating a picture from a written description. Results can be striking but often slip on details such as hands, text within the image and exact counts of objects. Rights and disclosure questions apply before you use the images commercially.
Text-to-speech: Turning written text into spoken audio. Modern systems produce natural-sounding voices with controllable tone and pace, and some can imitate a specific voice from a short sample, which raises consent and fraud risks.
Token: The unit of text a language model reads and writes: a whole word, part of a word, or punctuation. In English a token averages roughly three-quarters of a word, though it varies by model and language. Limits and API charges are usually counted in tokens.
Training: The process of adjusting a model's internal numbers so its outputs get closer to the examples it is shown. Training a large model from scratch takes huge amounts of data and computing power; most organisations use a model someone else trained.
Training data: The examples a model learns from. For large language models this is a vast mix of text such as public web pages, books and code, plus data curated or written for later training stages. The model's knowledge, gaps and biases all trace back to it.
Turing test: A test proposed by Alan Turing in his 1950 paper 'Computing Machinery and Intelligence': can a machine hold a text conversation well enough that a person cannot reliably tell it from a human? Passing it shows convincing conversation, not understanding or reliability.

Building with AI

API (application programming interface): A way for software to talk to other software. AI providers offer APIs so developers can send prompts and receive outputs from their own code, usually charged per token and with different data terms from the consumer chat app.
API key: A secret string that identifies your account when your code calls an API, and that the bill is charged to. Treat it like a password: never paste it into front-end code, public repositories or shared documents, and rotate it if it leaks.
Agent: An AI system that works towards a goal over several steps, deciding which tools to use and acting on the results with limited supervision. Definitions vary widely, and the word is used in marketing for simple chatbots too. The more an agent can do, the more its permissions and checks matter.
Agentic workflow: A process where an AI model decides some of the steps itself, rather than following a fixed sequence. It handles messy, varied tasks better than a fixed flow but is harder to predict, test and audit. Many business processes work better as a fixed workflow with AI at specific steps.
Batch processing: Sending many requests together to be processed in the background over a longer window, rather than waiting for each in real time. Providers that offer it often charge less, which suits jobs such as classifying a backlog of documents overnight.
Benchmark: A standard public test used to compare models, such as sets of exam questions, coding problems or maths puzzles. Useful for broad comparison, but scores rarely predict performance on your specific task, and popular benchmarks saturate as models improve.
Benchmark contamination: When benchmark questions or answers have leaked into a model's training data, so a high score partly reflects memory rather than ability. It is one reason to test models on your own unpublished examples.
CLAUDE.mdClaude Code: A Markdown file in a project that Claude Code reads for standing instructions, such as build commands, code conventions and things to avoid. It saves re-explaining the project each session; keep it short and current.
Chunking: Splitting documents into smaller passages before creating embeddings, so retrieval can return just the relevant part. Chunk size and boundaries matter: too small loses context, too large dilutes relevance. Poor chunking is a common cause of poor RAG answers.
Coding agent: An AI tool that can read a codebase, plan changes, edit several files, run commands and tests, and iterate on errors. It speeds up development but can introduce subtle bugs and security flaws, so its changes need review like any colleague's.
Computer use: An agent capability where the model sees screenshots and controls a mouse and keyboard to operate ordinary software or websites. It can handle tools without an API, but it is slower and more error-prone than direct integrations, and needs close supervision.
Context engineering: Deciding what information goes into a model's context window for each step: instructions, retrieved documents, tool results, memory and conversation history. As applications get more complex, what you give the model matters as much as how you phrase the request.
Eval (evaluation): A repeatable test of how well a model or AI feature performs on your task, using a set of inputs and a way to score the outputs. Evals are how you know whether a prompt change, a new model or a fix actually helped, rather than guessing from a few examples.
Fine-tuning: Further training an existing model on your own examples to adjust its behaviour, style or format for a specific task. It is good for consistency but poor at adding facts that change; for knowledge, RAG is usually cheaper and easier to keep current.
Golden dataset: A carefully chosen set of example inputs with agreed good answers, used to run evals. It should include ordinary cases, edge cases and past failures. It needs maintaining as the real work changes.
Hybrid search: Combining keyword search and semantic search, then merging the results. It catches both exact matches such as names and codes and matches by meaning, and is a common default for RAG systems.
LLM-as-judge: Using a language model to grade another model's outputs against criteria, so evals can run at scale. Quicker than human review, but judges have their own biases, such as favouring longer answers, so check a sample against human judgement.
LLMOps: The operational discipline of running language model applications reliably: versioning prompts and models, running evals before release, monitoring quality and cost, and managing incidents. It extends the older idea of MLOps to generative AI.
Latency: How long you wait for a response. For AI features it depends on model size, prompt length, output length, reasoning effort and server load. Users notice latency more than small differences in quality, so measure it early.
LoRA (low-rank adaptation): A cheap way to fine-tune a model by training a small set of added weights while leaving the original model untouched. The result is a small file that can be swapped in and out, widely used for customising open-weights language and image models.
MCP (Model Context Protocol): An open standard, introduced by Anthropic in November 2024, for connecting AI applications to external tools and data through small programs called MCP servers. One server can then work with any AI tool that supports the protocol. Connecting a server grants real access, so vet it as you would any integration.
Model routing: Sending each request to the most suitable model, such as a small fast model for simple questions and a larger one for hard ones. Done well it cuts cost and latency; it adds a component that can itself make mistakes.
Multi-agent system: A setup where several AI agents with different roles work together, such as one planning, several researching and one reviewing. It can help with large tasks but multiplies cost, and errors can pass between agents unnoticed.
No-code automation: Building workflows that connect apps and AI steps through a visual editor rather than writing code. It lets non-developers automate real processes quickly. It still needs testing, error handling and an owner, or it fails quietly.
Observability: Logging and tracing what an AI application actually does in production: prompts, retrieved context, tool calls, outputs, costs, latency and user feedback. Without it you cannot diagnose failures or prove what happened. Log with privacy in mind.
Open weights: A model whose trained weights are published so anyone can download, run and adapt it, usually under a licence with conditions. Open weights is not the same as open source: the training data and code are often not released.
Open-source AI: AI released with the freedom to use, study, modify and share it. The Open Source Initiative published a definition in 2024 that expects the weights, the code and enough information about the training data to be available. Many models marketed as open do not meet it.
Orchestration: The code or platform that coordinates the steps of an AI application: calling models, running tools, passing results along, handling errors and retries. Good orchestration is often what separates a demo from a dependable product.
Prompt caching: Reusing the processing of a repeated prompt prefix, such as a long system prompt or reference document, across requests. Where a provider supports it, it can cut both cost and latency substantially for applications that send the same context repeatedly.
RAG (retrieval-augmented generation): A pattern where the system first retrieves relevant passages from your own documents or data, then gives them to the model to answer from. It keeps answers current and specific without retraining, and allows citations. Answer quality depends heavily on the retrieval step.
Rate limit: A cap on how many requests or tokens you can send to an API in a given period. Exceed it and requests are rejected, typically with an HTTP 429 error. Production systems need to queue, retry with back-off, or spread load.
Reranking: A second pass that reorders retrieved passages by how well they actually answer the question, using a more precise model than the first search. It often improves RAG quality noticeably for little extra cost.
SDK (software development kit): A library provided for a programming language, such as Python or TypeScript, that wraps an API so you can call it with a few lines of code rather than building raw web requests. It usually handles retries, streaming and error types for you.
Sandbox (code execution): An isolated environment where AI-generated code or agent actions can run without touching real systems or data. Sandboxing limits the damage if the model does something wrong or is manipulated.
ScenarioMake: Make's name for an automated workflow, built visually as a chain of connected modules with routes and filters. Scenarios can include AI steps; test them with realistic data before switching them on.
Self-hosting: Running a model on your own servers or private cloud rather than calling a provider's API. It gives more control over data and costs at high volume, but you take on hardware, updates, security and monitoring, and the models you can run may be less capable.
Semantic search: Search by meaning rather than exact words, using embeddings, so 'staff holiday policy' can find a document titled 'annual leave entitlement'. It can miss exact identifiers such as product codes, which keyword search handles better.
SpacesHugging Face: Hugging Face's hosting for small AI demo apps, often built with Python frameworks such as Gradio or Streamlit. A quick way to try open models in the browser or share a prototype; check the licence of any model before commercial use.
Streaming: Sending a model's output to the user piece by piece as it is generated, rather than all at once at the end. It is why chat replies appear word by word. It improves perceived speed but complicates checking output before the user sees it.
Structured output: Getting a model to return data in a fixed format, usually JSON matching a schema you define, so your code can use it directly. Many APIs can enforce the format; the values inside it still need validating.
Throughput: How much work a system can handle over time, such as tokens per second or requests per minute. Latency is about one user's wait; throughput is about serving many users at once. Improving one can worsen the other.
Time to first token: How long before the first piece of a response appears. With streaming, a short time to first token makes a slow response feel fast, because people can start reading while the rest is generated.
Tool calling (function calling): Letting a model request that your software run a defined function, such as look up an order, search a database or send an email, and then use the result. The model only asks; your code decides whether to run it, which is where permissions and checks belong.
Vector database: A database designed to store embeddings and quickly find the ones closest in meaning to a query. It is the usual retrieval layer in RAG systems. Many general-purpose databases now offer vector search too.
Vibe coding: Building software by describing what you want to an AI tool and accepting its code largely without reading it. The phrase was coined in early 2025. Fine for prototypes and personal tools; risky for anything handling customer data or money unless someone reviews the code.
Webhook: A way for one app to notify another automatically when something happens, by sending data to a URL. Webhooks commonly trigger automations, such as starting an AI summary when a form is submitted.
ZapZapier: Zapier's name for an automated workflow: a trigger in one app followed by one or more actions, which can include AI steps. Zaps run unattended, so build in checks before anything is sent to customers.

Business and strategy

AI champion: A member of staff in a team who helps colleagues adopt AI, shares useful prompts and feeds back problems. A champion network spreads practical skills faster than central training alone.
AI governance: The policies, roles, processes and records an organisation uses to decide which AI to use, how, and who is accountable. Proportionate governance speeds adoption by giving people clear permission; heavy governance for low-risk uses drives people to shadow AI.
AI maturity: How far an organisation has progressed in using AI well, from ad hoc individual use to governed, measured use embedded in core processes. Maturity models are useful for spotting gaps; the stages are conventions, not standards.
AI register (inventory): A record of the AI systems an organisation uses or builds, with each one's owner, purpose, data, risk level and review date. It is the foundation for governance and for answering regulators, auditors and customers.
AI strategy: An organisation's plan for where AI will create value, what it will not be used for, and the data, skills, governance and budget needed. A good one starts from business problems, not from tools.
Acceptable use policy: A short, practical document telling staff which AI tools they may use, what data may go into them, what must be checked by a person, and what must be disclosed. The best ones fit on a page and are backed by approved tools.
Augmentation vs automation: Augmentation uses AI to help people do their work better or faster; automation hands the task to AI entirely. Most successful early uses are augmentation, because a person remains accountable for the result.
Build vs buy: The decision whether to create your own AI solution, buy a ready-made product, or configure a platform somewhere in between. Buy for common tasks; build when the process is a genuine differentiator and you can maintain it.
Centre of excellence: A small central team that sets AI standards, shares good practice, runs shared platforms and supports business units, rather than doing all AI work itself. It works best when it enables teams rather than acting as a gatekeeper.
Change management: The work of helping people adopt a new way of working: communication, training, redesigning roles and processes, and addressing worries honestly. Most AI programmes that stall do so here rather than on the technology.
Model-agnostic: Designed to work with models from different providers, so you can switch as prices, capabilities or terms change. It adds a little engineering effort and is usually worth it for systems expected to last.
Pilot: A limited real-world trial of an AI solution with actual users and data, designed to measure value and risk before wider rollout. Set success criteria and a decision date before it starts.
Pilot purgatory: When an organisation runs many AI pilots but few ever reach full production, because no one set success criteria, funded the rollout or owned the change. The fix is fewer pilots, each with a clear path to scale or to stop.
Proof of concept: A small, quick build to test whether an AI idea can work at all, usually with sample data. Success shows feasibility, not value at scale: real data, edge cases, integration and adoption are where most of the effort lies.
ROI (return on investment): The benefit from an investment relative to its cost. For AI, measure outcomes such as time saved and actually redeployed, error rates, revenue or customer satisfaction, against a baseline taken before rollout. Time saved that is not used for something else is not a return.
Responsible AI: An umbrella term for developing and using AI in ways that are fair, safe, transparent, accountable and lawful. It is only meaningful when turned into specific practices, such as testing for bias, documenting decisions and giving people a route to challenge outcomes.
SLA (service level agreement): The part of a supplier contract that commits to measurable service levels, such as uptime and support response times, often with credits if they are missed. Check whether an AI provider's SLA covers the API or plan you depend on.
Seat-based vs usage-based pricing: Two common ways AI is charged: a fixed fee per user, or a charge for what you consume, such as tokens, credits or tasks. Seats are predictable but waste money on light users; usage scales with value but can surprise you. Many tools mix both.
TCO (total cost of ownership): The full cost of an AI solution over its life, not just licences or tokens: integration, data preparation, security review, training, monitoring, evals, support, and the staff time to run it. TCO often decides build-versus-buy questions.
Unit economics (cost per task): What it costs to complete one unit of work with AI, such as one support ticket or one document processed, including model costs, retries and human review. Comparing it with the current cost per task gives a more honest business case than headline productivity claims.
Use case: A specific task or process where AI is applied, such as drafting responses to customer emails or extracting data from invoices. Strong use cases have a clear owner, measurable value, available data and tolerable consequences when the AI is wrong.
Vendor lock-in: Dependence on one AI provider that makes switching costly, through proprietary features, stored data, custom integrations or staff habits. Reduce it by keeping prompts, data and evals portable and by separating your application logic from any single model.

Data and privacy

Anonymisation: Processing data so people can no longer be identified, by any means reasonably likely to be used. Truly anonymous data falls outside UK GDPR, but it is harder to achieve than it looks: removing names is rarely enough.
Data classification: Labelling information by sensitivity, such as public, internal, confidential and restricted, so people know what can go into which tools. A simple classification scheme is one of the most useful parts of an AI acceptable use policy.
Data labelling: Adding the correct answers or categories to raw data so it can be used for supervised learning or evaluation, often done by people. Labelling quality, consistency and the working conditions of labellers all affect the final model.
Data leakage: Sensitive information ending up somewhere it should not, such as confidential documents pasted into a consumer AI tool, or an AI feature revealing one user's data to another. It is one of the most common real-world AI risks for organisations.
Data minimisation: The principle of using only the personal data you actually need for the purpose. In practice: strip names and irrelevant details before pasting material into an AI tool, and do not upload whole datasets when a sample will do.
Data processing agreement: The contract between an organisation and a supplier that processes personal data on its behalf, setting out what the supplier may do with it. UK GDPR requires one; business and API AI plans usually offer one, and consumer plans generally do not.
Data provenance: A record of where data came from, how it was collected and what has been done to it. It underpins decisions about lawful use, copyright, quality and bias in AI systems, and is increasingly asked for in procurement and regulation.
Data residency: Where data is physically stored and processed, such as in the UK, the EU or the US. Some organisations and contracts require particular locations. Storage location and processing location can differ, so ask about both.
Data retention: How long a tool or provider keeps your prompts, files and outputs, and for what purposes. Retention periods differ between consumer, business and API offerings, and can change, so check the current terms for the plan you actually use.
PII (personally identifiable information): A mainly American term for information that identifies someone, often used in AI product documentation. UK law uses the broader concept of personal data, which can include information that identifies someone only in combination with other data.
Personal data: Any information relating to an identified or identifiable living person, such as a name, email address, location, or an ID number. Pasting it into an AI tool is processing under UK GDPR, so you need a lawful basis and a tool whose terms fit.
Pseudonymisation: Replacing identifying details with codes, while keeping a separate key that could reverse it. It reduces risk and is good practice before using data with AI, but pseudonymised data is still personal data under UK GDPR.
Special category data: Personal data that UK GDPR gives extra protection: health, racial or ethnic origin, religious or philosophical beliefs, political opinions, trade union membership, genetic and biometric data used for identification, sex life and sexual orientation. Keep it out of AI tools unless you have a clear legal basis and safeguards.
Sub-processor: A third party your supplier uses to process your data, such as the cloud host behind an AI tool or the model provider behind an AI feature in other software. Suppliers should list their sub-processors; check where your data actually goes.
Synthetic data: Artificially generated data that mimics the patterns of real data. Used to train or test models when real data is scarce or sensitive. It can reduce privacy risk, but can also leak details of the original data or miss the rare cases that matter most.
Training opt-out: A setting or contract term that stops a provider using your conversations to train future models. Defaults differ between tools and between consumer and business plans, and have changed over time, so check the data controls in each tool you use.
Zero data retention: An arrangement where the provider processes your requests without storing prompts or outputs afterwards, apart from what is needed to serve them. Usually available only to eligible business or API customers by agreement, and some features may be unavailable under it.

How models work

Attention: The mechanism in a transformer that lets the model weigh how relevant each token is to every other token when working out meaning. It is how the model links 'it' back to the right noun several sentences earlier.
Backpropagation: The technique that calculates how much each weight in a neural network contributed to an error, working backwards from the output layer. Those calculations tell gradient descent how to adjust each weight.
Base model: A model straight out of pretraining, before it has been tuned to follow instructions or hold a conversation. Given a question it may continue with more questions rather than answer. Developers sometimes use base models as a starting point for their own fine-tuning.
Chain of thought: Step-by-step reasoning written out before a final answer, either prompted ('think it through step by step') or built into reasoning models. It often improves accuracy on complex tasks. Research suggests the visible reasoning does not always faithfully reflect how the model reached its answer.
Compute: Shorthand for the processing power used to train and run AI models, usually on large clusters of specialised chips in data centres. It is one of the three main inputs to AI capability, alongside data and algorithms, and carries significant energy and cost.
Constitutional AI: A training approach, published by Anthropic in 2022, in which a model critiques and revises its own responses against a written set of principles, and AI-generated feedback partly replaces human ratings. The aim is to make the values guiding behaviour more explicit.
Diffusion model: The main approach behind modern image and video generators. It learns to reverse a process of gradually adding noise to images, so it can start from pure noise and remove it step by step into a picture that matches the prompt.
Distillation: Training a smaller 'student' model to imitate the outputs of a larger 'teacher' model. The student is cheaper and faster, and keeps much of the teacher's ability on the tasks it was distilled for, though usually not all of it.
Embedding: A list of numbers that represents the meaning of a piece of text, an image or other data, so that similar things end up with similar numbers. Embeddings let software find related content by meaning rather than exact words, which underpins semantic search and RAG.
Emergent abilities: Capabilities that seem to appear suddenly once a model passes a certain size, rather than improving gradually. The idea is contested: some researchers argue the sudden jumps partly reflect how the abilities were measured.
Foundation model: A large model trained on broad data that can be adapted to many tasks, such as a large language model that underpins chat, search and coding tools. The term highlights that many products are built on a small number of such models.
Frontier model: A loose label for the most capable models available at a given moment. Because the frontier keeps moving, a model described this way today will not be one in a year or two. Policy documents sometimes use the term with a more specific meaning.
GAN (generative adversarial network): A design, introduced in 2014, where two networks compete: one generates fake examples and the other tries to spot them, and both improve. GANs produced the first convincing synthetic faces; diffusion models have since overtaken them for most image generation.
GPU (graphics processing unit): A chip originally designed for rendering graphics whose ability to do many calculations in parallel suits the maths of training and running neural networks. Access to GPUs is a major constraint and cost in AI.
Gradient descent: The optimisation method used to train neural networks. It works out which direction to adjust each weight to reduce the loss a little, takes a small step, and repeats millions of times.
Inference: Running a trained model to get an output, as opposed to training it. Every prompt you send triggers inference. For organisations using AI at scale, inference is where most of the ongoing cost sits.
Instruction tuning: Further training on examples of instructions paired with good responses, which turns a base model into one that follows requests. It is one of the steps that makes a model feel like an assistant.
Latent space: The internal mathematical space in which a model represents data, where similar concepts sit close together. Image generators work largely in latent space, which is why nudging a prompt can shift style or content smoothly.
Loss function: The measure of how wrong a model's outputs are during training. Training works by repeatedly nudging the weights to make this number smaller. What you choose to measure shapes what the model learns to be good at.
Mixture of experts: A model design split into many specialist sub-networks, with a router that sends each token to only a few of them. The model can hold a very large number of parameters while using only a fraction for each token, which makes it cheaper to run than its total size suggests.
Multimodal: Able to take in or produce more than one kind of data, such as text, images, audio and video. A multimodal assistant can read a photo of a receipt or answer a spoken question. Understanding of each mode can be uneven, so test the one you rely on.
Next-token prediction: The core task a language model is trained on: given the text so far, predict the most likely next token. Generation repeats that prediction one token at a time. Everything else these models do, from answering questions to writing code, is built on this.
Non-determinism: The fact that the same prompt can produce different outputs each time. It comes from sampling and from how computations run on hardware. It matters for testing: one good answer does not prove the next will be good too.
Overfitting: When a model learns its training examples too closely, including their noise and quirks, and so performs well on them but poorly on new data. It is why models are always tested on data they have not seen.
Parameters: The learned numbers inside a model, mainly its weights, that training adjusts. Large language models have billions of them. More parameters can mean more capability, but training data, training method and design matter just as much, so parameter count alone is a poor guide to quality.
Pretraining: The first and most expensive stage of building a large language model, where it learns general language and knowledge by predicting text across a huge dataset. The result is a base model that continues text well but does not yet follow instructions reliably.
Quantisation: Storing a model's weights with fewer bits, for example 8 or 4 instead of 16, so it needs less memory and runs faster, often on smaller hardware. It usually costs some quality, more so at lower bit widths.
RLHF (reinforcement learning from human feedback): A training method where people compare or rate model responses, a separate reward model learns their preferences, and the language model is then trained to score well against it. It makes models more helpful and polite, and can also make them more eager to please.
Reasoning model: A language model trained to work through a problem in intermediate steps before answering, often called thinking. It tends to do better on maths, logic, coding and multi-step planning, at the cost of slower and more expensive responses. It can still reason its way to a wrong answer.
Reinforcement learning: Training by trial and reward: the system takes actions, receives a score, and gradually learns which actions lead to better scores. It is used in game-playing AI, robotics and, increasingly, to train language models to reason through problems.
Scaling laws: Observed relationships showing that model performance improves predictably as training compute, data and parameters increase together. They guided the push towards ever larger models. They describe a trend in measured loss, not a guarantee of any particular capability.
Self-supervised learning: Training where the labels come from the data itself, for example hiding the next word and asking the model to predict it. It lets models learn from vast unlabelled datasets and is how large language models are pretrained.
Small language model: A language model with relatively few parameters, designed to run cheaply, quickly or on a laptop or phone. Well suited to narrow tasks such as classification or extraction, but weaker at broad knowledge and complex reasoning than larger models.
Supervised learning: Training on examples that come with the right answer attached, such as photos labelled 'cat' or 'dog', or loan applications labelled 'repaid' or 'defaulted'. The model learns to map inputs to those labels. Label quality sets a ceiling on model quality.
Temperature: A setting that controls how random a model's word choices are. Low temperature makes output more predictable and repetitive; high temperature makes it more varied and more likely to drift. It is usually adjustable through an API or developer console rather than in consumer chat apps.
Test-time compute: Computing power spent while answering a question rather than during training, for example by letting a model reason for longer or try several approaches. More test-time compute often buys better answers on hard problems, at the cost of time and money.
Tokeniser: The component that splits text into tokens before a model sees it and joins tokens back into text afterwards. Tokenisation explains some odd weaknesses, such as models struggling to count the letters in a word, because they see chunks rather than individual characters.
Top-p (nucleus sampling): A setting that limits the model's choice of next token to the smallest set of options whose combined probability reaches p, such as 0.9. Like temperature, it trades predictability for variety. Most guidance suggests adjusting one or the other, not both.
Transformer: The neural network design behind nearly all modern language models, introduced by Google researchers in the 2017 paper 'Attention Is All You Need'. Its attention mechanism lets it process a whole passage in parallel and relate every word to every other, which made training at huge scale practical.
Unsupervised learning: Training on data without labels, where the model finds structure by itself, such as grouping customers with similar buying habits. Useful for exploration, but the groups it finds still need a person to judge whether they mean anything.
Vision-language model: A model that can reason about images and text together, for example describing a chart, reading a scanned form or answering questions about a photo. It may misread small text, dense tables and precise measurements.
Weights: The numerical values on the connections in a neural network that determine how strongly one unit influences the next. A model's weights are, in effect, the model: releasing them lets anyone run it.

Regulation

Bletchley Declaration: A statement agreed at the AI Safety Summit hosted by the UK at Bletchley Park on 1 and 2 November 2023, in which countries including the US and China, and the EU, recognised risks from frontier AI and committed to international cooperation on AI safety.
Computer-generated works: A category in UK copyright law for works generated by computer with no human author, where the author is treated as the person who made the arrangements for their creation. How it applies to generative AI output is untested and has been under government review.
Controller: Under UK GDPR, the organisation that decides why and how personal data is processed, and carries primary responsibility for compliance. Using an AI supplier does not transfer that responsibility to them.
Council of Europe AI Convention: The Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law, the first legally binding international treaty on AI, opened for signature on 5 September 2024. The UK signed it that day; it binds each country only once that country has ratified it.
Data (Use and Access) Act 2025: A UK Act that received Royal Assent on 19 June 2025 and amends UK GDPR and related law, including the rules on automated decision-making and the structure of the data protection regulator. Its provisions commence in stages, so check which are in force.
Data Protection Act 2018: The UK statute that sits alongside UK GDPR, adding exemptions, rules on law enforcement and intelligence processing, and the powers of the data protection regulator. Most practical AI questions are answered by reading the two together.
Data protection by design and default: The UK GDPR requirement to build privacy into systems from the start and to process only the personal data needed by default. For AI projects it means deciding data flows, retention and access before building, not after.
EU AI Act: Regulation (EU) 2024/1689, the EU's risk-based law on AI, which entered into force on 1 August 2024 with obligations applying in stages. It can apply to UK organisations that place AI systems on the EU market or whose AI outputs are used in the EU.
Equality Act 2010: The main anti-discrimination law in Great Britain, protecting characteristics such as age, disability, race, religion or belief, sex and sexual orientation. It applies to decisions made with AI just as to decisions made by people, including in recruitment and service provision.
Lawful basis: One of the six grounds UK GDPR requires for processing personal data: consent, contract, legal obligation, vital interests, public task or legitimate interests. You need one for each purpose, including using personal data with an AI tool, and should record which you rely on.
Legitimate interests: The most flexible lawful basis under UK GDPR, used when processing is necessary for a genuine interest and does not override the person's rights. It requires a documented three-part test covering purpose, necessity and balance, and is often relied on for AI uses.
Online Safety Act 2023: UK law, regulated by Ofcom, placing duties on user-to-user services and search services to tackle illegal content and protect children. Ofcom has said that generative AI tools can fall within scope, for example where users can share AI-generated content with each other.
Processor: Under UK GDPR, an organisation that processes personal data on a controller's behalf and under its instructions, such as an AI provider on a business contract. If a provider uses your data for its own purposes, such as training, it may be acting as a controller for that use.
Subject access request: A person's right under UK GDPR to obtain a copy of the personal data an organisation holds about them. Organisations normally have one month to respond. Records and outputs from AI systems that contain someone's personal data can fall within scope.
Text and data mining exception: The UK copyright exception, in section 29A of the Copyright, Designs and Patents Act 1988, that allows computational analysis of lawfully accessed works for non-commercial research only. Commercial AI training is not covered. The government has consulted on changing the law, so check the current position.
Transparency obligations (AI Act Article 50): EU AI Act duties, including to tell people when they are interacting with an AI system, to mark synthetic content in a machine-readable way, and to disclose deepfakes and certain AI-generated text published to inform the public. They apply alongside, not instead of, other obligations.
UK GDPR: The UK's version of the General Data Protection Regulation, retained after Brexit and read alongside the Data Protection Act 2018. It governs any use of personal data in AI, from pasting a customer email into a chatbot to training a model.

Risk and safety

AI incident: An event where an AI system causes or nearly causes harm, such as a chatbot giving customers wrong refund terms or an agent deleting data. Treat AI incidents like other operational incidents: log them, investigate the cause and share what you learn.
AI safety: The field concerned with preventing harm from AI systems, from everyday failures such as biased or false outputs to serious risks from misuse or loss of control of highly capable systems. People use the term with very different scopes, so ask which is meant.
AI text detector: A tool that claims to tell whether text was written by AI. They are unreliable: they produce false positives, particularly on writing by non-native English speakers, and are easily fooled by light editing. Do not use one as the sole basis for an accusation.
Adversarial example: An input subtly altered to fool a model, such as an image with changes invisible to people that makes a classifier see something else. It shows that models can fail in ways humans would never predict.
Alignment: Making an AI system's behaviour match the intentions and values of the people it serves, including in situations its designers did not anticipate. It covers everything from following instructions faithfully to avoiding harmful goals in more capable systems.
Automation bias: The human tendency to over-trust an automated system's output and stop checking it, especially when it is usually right. It undermines human review: a reviewer who approves everything is not a safeguard.
Bias: Systematic unfairness in an AI system's outputs, such as a CV screener that favours one group or an image generator that stereotypes professions. It usually comes from skewed training data, design choices or how the system is used. It has to be tested for; it does not show up on its own.
Content credentials (C2PA): A technical standard from the Coalition for Content Provenance and Authenticity that attaches signed information to a file showing how it was made and edited, including any AI involvement. Useful where it is preserved, but the metadata can be stripped when files are re-saved or uploaded.
Data poisoning: Deliberately corrupting a model's training or retrieval data so it behaves badly, for example planting content that teaches a hidden trigger or skews answers. It is a risk for anyone training models or feeding them content from sources they do not control.
Deepfake: AI-generated or AI-altered image, video or audio that makes a real person appear to say or do something they did not. Used in fraud, including fake calls from 'the boss' asking for urgent payments. Verify unusual requests through a separate, known channel.
Drift: Gradual decline in an AI system's performance because the real world changes (data drift) or because the model or provider behaviour changes underneath you (model drift). It is why AI systems need ongoing monitoring, not one-off testing.
Excessive agency: Giving an AI system more permissions, tools or autonomy than its task needs, so a mistake or manipulation can do real damage. It is one of the risks in the OWASP Top 10 for LLM applications. The fix is least privilege and approval for consequential actions.
Explainability: The ability to give a meaningful account of why an AI system produced a particular output, in terms the affected person can understand. It matters most for decisions about people, such as credit, hiring or benefits, and is hard to achieve with large models.
Fairness: Whether an AI system treats people and groups equitably. There are several mathematical definitions of fairness, and some cannot all be satisfied at once, so organisations must decide which matters for their context and justify it.
Guardrails: Controls around an AI system that keep it within acceptable behaviour, such as filters on inputs and outputs, topic restrictions, limits on actions and checks before anything is sent. Guardrails reduce risk; they do not remove it.
Human in the loop: A design where a person reviews and approves an AI system's output before it takes effect, such as checking a drafted letter before it is sent. It only works if the reviewer has the time, information and authority to say no.
Human on the loop: A design where the AI acts on its own but a person monitors it and can step in or stop it. Suited to high-volume, lower-risk tasks; it needs good alerts and a clear way to intervene.
Indirect prompt injection: Prompt injection hidden in content the AI reads on the user's behalf, such as an email, document, calendar invite or web page, rather than typed by the user. It is the main security risk for agents and assistants connected to email, files and the web.
Interpretability: Research into what is actually happening inside a model, for example which internal features represent which concepts. It aims to understand models from the inside rather than only judging their outputs, and remains an active and incomplete field.
Jailbreak: A prompt designed to get a model to break its own safety rules, for example through role-play or elaborate hypotheticals. Providers patch known jailbreaks, and new ones keep appearing, so model safeguards should not be your only line of defence.
Least privilege: The security principle of giving a user, system or AI agent only the access it needs for its task and nothing more, such as read-only access to one folder rather than the whole drive. It limits the damage when something goes wrong.
Model card: A document published with a model describing what it is for, how it was trained and evaluated, known limitations and risks, and where it should not be used. Read one before adopting a model; note what it leaves out as well as what it says.
Model collapse: The degradation research has shown can happen when models are trained repeatedly on content generated by other models: rare details disappear and outputs become narrower and less accurate. A reason human-created data and provenance remain valuable.
Prompt injection: An attack where text crafted to look like instructions tricks a model into ignoring its original instructions, such as a web page telling an AI assistant to reveal data. There is no complete fix yet, so limit what an AI system can access and do.
Red teaming: Deliberately attacking an AI system to find its weaknesses before others do: trying to make it produce harmful content, leak data, show bias or take unintended actions. Useful for any AI feature exposed to customers or the public, not just for model developers.
Reward hacking: When a model finds a way to score well on its training objective without doing what was intended, such as a coding model making tests pass by editing the tests. It shows how hard it is to specify exactly what you want.
Safety classifier: A separate model that screens inputs or outputs for categories such as hate, self-harm or personal data, and blocks or flags them. Classifiers can over-block harmless content and miss harmful content phrased in new ways, so tune them to your context.
Shadow AI: Staff using AI tools the organisation has not approved or does not know about, often personal accounts on consumer plans. It is usually a sign that people need AI for their work and lack a sanctioned option. Banning it rarely works; providing approved tools and clear rules does better.
Sycophancy: A model's tendency to agree with the user, praise their ideas or change a correct answer when pushed back on. It is partly a side effect of training on human approval. Ask for criticism explicitly, and be wary when a model instantly agrees it was wrong.
System card: A developer's detailed report on a model or AI system as released, typically covering safety testing, red-team findings, risk assessments and mitigations. Longer than a model card; the most useful sections are usually the known limitations.
Voice cloning: Creating a synthetic copy of a specific person's voice from recordings. It has legitimate uses in accessibility and production, but also enables impersonation scams. Get clear consent before cloning anyone's voice, including your own staff's.
Watermarking: Embedding a hidden, machine-detectable signal in AI-generated content to mark its origin. It helps with identification, but watermarks can often be weakened by editing, and there is no universal detector across providers.

Using AI

AI search (answer engine): Search that answers your question in prose, built from web results, rather than returning a list of links. Faster for many questions, but the summary can misread or misquote its sources, so open the citations for anything that matters.
ArtifactsClaude: Claude's panel for substantial, standalone output such as documents, code, diagrams or small interactive apps, shown alongside the chat so you can refine it and reuse it. The spelling is the product's own.
Citations: Links or references an AI tool attaches to show where a claim came from. Their presence is not proof: a cited page may not say what the summary claims. Spot-check citations, especially numbers and quotes.
Custom GPTChatGPT: A version of ChatGPT configured with its own instructions, knowledge files and optional capabilities or connections to other services, for a specific task. It can be kept private, shared by link or published for others to use.
Custom instructions: Standing preferences you set once in an assistant's settings, such as your role, preferred tone or format, which it then applies to every new conversation. Names and options vary by tool. Review them occasionally; stale instructions quietly skew results.
Deep research: A mode in several AI tools where the assistant runs many searches, reads the results and writes a long report with citations, taking minutes rather than seconds. Useful as a starting point; the sources and the conclusions drawn from them still need checking.
Delimiters: Markers such as triple quotes, headings or XML-style tags that separate your instructions from the material you want processed. They help the model tell the task apart from the text, and reduce the chance that instructions hidden in a pasted document are followed.
Few-shot prompting: Including a handful of worked examples in your prompt so the model copies the pattern, such as three product descriptions in your house style before asking for a fourth. The model may copy the examples too closely, so vary them.
GemGemini: A customised version of Gemini that you set up with your own instructions, and optionally files, for a task you repeat, such as a writing coach or a meeting-notes formatter. It saves re-explaining the same brief each time.
Grounding: Tying a model's answer to specific source material, such as web results, your documents or a database, rather than relying on what it absorbed in training. Grounding reduces hallucination but does not remove it; the model can still misread the source.
Inpainting: Editing part of an image with AI by selecting an area and describing what should replace it, such as removing a stray object or changing a jacket's colour. Its counterpart, outpainting, extends an image beyond its original edges.
Iterative prompting: Improving a result through follow-up messages rather than expecting a perfect answer first time: 'shorter', 'more formal', 'use UK spelling', 'add a table'. Often faster than trying to write the perfect prompt up front.
Memory: A feature in some assistants that stores facts about you across conversations and uses them later. It saves repetition, but it can also carry outdated or wrong assumptions forward. Check what is stored and how to edit or delete it.
Negative prompt: In image and video generators, a list of things you do not want to appear, such as 'text, watermark, extra fingers'. Not every tool supports one; where it does not, describing what you want positively usually works better than saying 'no X' in the main prompt.
Project (workspace): A space in some assistants that groups related chats with shared files and instructions, so each new conversation starts with the same background. Useful for ongoing work such as a report or a client account.
Prompt engineering: The practice of designing prompts to get reliable, useful output: stating the task, giving context and examples, setting constraints and specifying the format. Less about magic phrases than about writing a clear brief, as you would for a new colleague.
Prompt template: A reusable prompt with blanks to fill in, for a task you do repeatedly, such as summarising a meeting or drafting a job advert. Templates make results more consistent across a team and are easy to improve once shared.
Role prompting: Asking the model to take on a role, such as 'act as an experienced HR adviser'. It can shift tone and focus usefully, but it does not give the model expertise it lacks, and it does not make the answer more accurate on its own.
Seed: A number that sets the starting randomness for an image or other generation. Reusing the same seed with the same prompt and settings lets you reproduce or make small controlled changes to a result, where the tool exposes it.
Summarisation: One of the most common and useful AI tasks: condensing a document, thread or meeting into key points. Summaries can drop important caveats or overstate certainty, so for decisions or anything sent to others, check against the original.
System prompt: Instructions given to a model before the conversation starts, setting its role, tone, rules and limits. Products use hidden system prompts; developers write their own. Users can often override or extract them, so do not treat a system prompt as a security control.
Temporary chat: A conversation mode in some assistants that is not saved to your history or used to personalise future chats. It is not the same as a guarantee that nothing is retained; read the tool's own explanation of what is kept and for how long.
Transcription: Turning recorded speech into text, often with speaker labels and an AI summary. Tell people when you are recording and transcribing, and check names, numbers and action points before relying on them.
Upscaling: Increasing an image's resolution with AI, which invents plausible fine detail rather than recovering real detail. Fine for design work; not reliable where the detail is evidence, such as reading a number plate.
Voice mode: Talking to an assistant out loud and hearing it reply, often in real time with the option to interrupt. Handy on the move or for practising conversations. Spoken answers are harder to check than text, so ask for a written summary when accuracy matters.
Zero-shot prompting: Asking a model to do a task with no examples, just the instruction. It works well for common tasks. When the output format or style matters, adding examples usually helps.

You know the words. Now learn to use AI well.

The free placement assessment builds you a personal path through our 18 released modules, from your first prompt to running AI across an organisation.

Find your level