AI & Digital Skills
AI buzzword dictionary
The AI words you keep hearing, explained without jargon. Each entry gives a short definition, what it really means, an everyday example and anything to be careful about.
AI agent
An AI system that can take a series of actions towards a goal, not just answer a single question.
What this actually means
An agent can plan steps and use tools — searching the web, filling in forms, sending messages or running code — then decide what to do next based on the results.
The more an agent can do on its own, the more important it is to limit what it has access to and to review its actions.
Example
You ask an agent to find three hotel options under £120 near a venue and add the best one to your calendar. It searches, compares and creates the calendar entry.
Where you may meet it
Booking assistants, coding assistants, customer-service systems and office automation.
API (application programming interface)
A defined way for one piece of software to request data or actions from another.
What this actually means
An API is like a menu and ordering system between programs: it lists what you can ask for and the format the answer comes back in.
Many AI tools are available through APIs, so other apps can add AI features without building their own models.
Example
A weather app shows tomorrow's forecast by asking a weather service's API, not by measuring the weather itself.
Where you may meet it
Anywhere apps connect: online payments, maps in websites, 'sign in with' buttons, AI features inside other software.
Artificial intelligence (AI)
Computer systems that perform tasks we normally associate with human thinking, such as recognising speech, sorting images or writing text.
What this actually means
AI is an umbrella term, not a single technology. It covers everything from the spam filter in your email to chat assistants that write whole paragraphs.
Most AI you meet today does not 'understand' things the way a person does. It finds patterns in large amounts of data and uses them to make predictions: which email is spam, which word comes next, which face matches a photo.
Example
Your phone's photo app groups pictures of the same person together. Nobody labelled every photo; a model learned which facial features tend to belong to the same face.
Where you may meet it
Search engines, sat-nav route planning, bank fraud alerts, voice assistants, customer-service chat, translation apps and writing tools.
Context window
The maximum amount of text a model can take into account at one time.
What this actually means
Everything in a conversation — your messages, the AI's replies and any documents you paste — has to fit inside the context window, measured in tokens.
When a conversation grows past that limit, earlier parts may be dropped or summarised, and the AI can appear to 'forget' them.
Example
You paste a long contract and ask about clause 40. If the document is too long for the window, the tool may miss the clause entirely.
Where you may meet it
Long chats, document summaries, and tools that advertise how many pages they can read.
Deepfake
Realistic fake video, audio or images of real people, made with AI.
What this actually means
Deepfakes can make it look or sound as though someone said or did something they never did.
They are used in scams, misinformation and harassment, as well as in legitimate film and accessibility work.
Example
A finance worker receives a voice message that sounds like their manager, urgently asking for a payment to a new supplier.
Where you may meet it
Social media, phone and video-call scams, fake celebrity endorsements in adverts.
Generative AI
AI that creates new content — text, images, audio, video or code — in response to an instruction.
What this actually means
Older AI mostly sorted or scored things. Generative AI produces something new: a draft email, a picture of a cat in a spacesuit, a summary of a report.
It builds its output from patterns learned during training. The result can be fluent and convincing while still containing mistakes.
Example
You type 'Write a polite reminder that an invoice is 14 days overdue' and receive a ready-to-edit email in seconds.
Where you may meet it
Chat assistants, image makers, 'help me write' buttons in email and office software, meeting-summary tools.
Hallucination
When an AI confidently produces information that is false or made up.
What this actually means
Because a language model generates likely-sounding text, it can invent facts, quotes, statistics or references that look genuine.
It is not lying on purpose; it has no built-in sense of truth. That is why checking matters.
Example
Asked for sources, an assistant lists a journal article with a real-sounding title and authors — but the article does not exist.
Where you may meet it
Anywhere AI gives facts: research summaries, legal or medical questions, citations, product specifications.
AI hallucination explained in full: causes, types and how to check →
Large language model (LLM)
A type of AI trained on huge amounts of text to predict and generate language.
What this actually means
An LLM works by repeatedly predicting a likely next piece of text, given everything before it. Done billions of times over, this produces answers, stories, code and summaries.
'Large' refers to the size of the model (billions of internal settings) and the volume of text it was trained on.
Example
When a chat assistant answers 'What is the capital of Australia?' with 'Canberra', it is producing the continuation that best fits the patterns it learned.
Where you may meet it
Behind most chat assistants and AI writing features.
Machine learning
A way of building AI where a computer learns patterns from examples rather than following hand-written rules.
What this actually means
Rather than programming every rule ('if the email mentions a lottery, mark it spam'), developers show the system many labelled examples and let it find the patterns.
Most modern AI, including language models, is built with machine learning.
Example
A bank trains a model on past transactions marked 'genuine' or 'fraud'. It then flags new transactions that look like the fraudulent ones.
Where you may meet it
Recommendations on streaming services, fraud detection, predictive text, medical image screening.
Multimodal AI
AI that can work with more than one type of input or output, such as text, images, audio and video.
What this actually means
A multimodal assistant might read a photo of a document, listen to a question, and reply in text or speech.
It widens what you can ask, but each type of input brings its own chances for error — a blurry photo or a noisy recording still causes mistakes.
Example
You photograph a meter reading and ask, 'What is the reading and how much has it increased since 45,210?'
Where you may meet it
Phone assistants, accessibility tools that describe images, apps that translate signs through the camera.
Prompt
The instruction or question you give an AI tool.
What this actually means
A prompt is everything you type (or attach) to tell the AI what you want. Clear prompts usually get more useful results.
Good prompts say who the output is for, what it should contain, the format, and any limits such as length or tone.
Example
Weak: 'Write about our sale.' Stronger: 'Write a 60-word email to existing customers announcing 20% off garden tools until Sunday. Friendly, no exclamation marks.'
Where you may meet it
Every chat assistant or AI feature that has a text box.
RAG (retrieval-augmented generation)
A method where an AI first looks up relevant documents, then uses them to write its answer.
What this actually means
Instead of relying only on what it learned in training, the system searches a chosen set of documents — a company handbook, a help centre — and passes the relevant passages to the model.
This usually makes answers more accurate and up to date, and lets the tool show where information came from.
Example
An HR assistant answers 'How many days of annual leave do I get?' by quoting the current staff handbook rather than guessing.
Where you may meet it
Workplace chatbots, 'ask your documents' features and support assistants.
Token
A small chunk of text — often part of a word — that language models read and write.
What this actually means
Models do not see whole words. They split text into tokens: common words may be one token, longer or rarer words several.
Tokens matter because tools measure limits and prices in them. As a rough guide in English, 100 tokens is about 75 words.
Example
'Unbelievable' might be split into 'Un', 'believ' and 'able' — three tokens for one word.
Where you may meet it
Pricing pages for AI services, usage limits, and error messages such as 'input too long'.
Training data
The examples used to teach an AI model — text, images, recordings or records.
What this actually means
A model's abilities and blind spots come largely from its training data. If a topic was rare in the data, the model will be weaker at it.
Training data has a cut-off date, so a model may not know about recent events unless it can search.
Example
A speech-recognition tool trained mostly on one accent may make more mistakes with other accents.
Where you may meet it
Privacy policies ('we may use your content to improve our models'), news about copyright and AI.
Last reviewed: 25 September 2026
