The plain-English AI glossary

An AI system that can take actions on your behalf — not just answer questions.

A packaged AI persona built to help with a specific role or workflow.

Systematic errors in AI output caused by imbalanced training data or design.

How prepared a business is to adopt AI — across data, people, process, and governance.

The field focused on preventing AI systems from causing harm — accidental or deliberate.

The problem of making AI do what humans actually want — safely and helpfully.

A doorway that lets one piece of software talk to another — how apps call AI models.

A hypothetical AI that matches or exceeds humans across essentially all cognitive tasks.

A theoretical AI that surpasses the best human minds across every domain.

The trick that lets a model weigh which parts of the input matter most when producing each word of output.

Asking the AI to reason step-by-step before giving its final answer.

A conversational program that answers user messages, often powered by an LLM.

The maximum amount of text an AI can hold in mind at once.

An AI that works alongside a human inside their existing tool — suggesting rather than replacing.

The AI architecture behind most image and video generators.

Training a small, fast model to mimic a large, slow one.

A list of numbers that represents the meaning of a piece of text.

A structured test of an AI's accuracy, safety, or usefulness on specific tasks.

Giving the AI a handful of examples in the prompt so it copies the pattern.

Training an existing model further on your own examples to change how it behaves.

A large, general-purpose model trained on broad data that can be adapted to many downstream tasks.

AI that creates new content — text, images, audio, video, or code — rather than only analysing existing data.

Connecting an AI's answers to verified sources so it can't just make things up.

Rules and filters that keep an AI safe, on-topic, and on-brand.

When an AI confidently makes something up.

The act of running a trained model to get an answer — as opposed to training it.

A prompt trick that gets an AI to bypass its safety rules.

A neural network trained on huge text collections to predict the next word — the engine behind ChatGPT, Claude and Gemini.

How long the AI takes to start (or finish) replying.

The ability for an AI to remember facts from earlier conversations.

A model architecture that routes each input to a small subset of specialised "expert" sub-networks.

An open standard from Anthropic that lets any AI app securely connect to any tool or data source.

A model that understands more than one type of input — e.g. text, images, and audio together.

Tools that let non-developers build AI applications with visual interfaces instead of code.

A model whose weights are freely downloadable, so anyone can run or modify it.

A defined personality, role, and voice given to an AI assistant.

The instruction you give an AI to tell it what to do.

The craft of writing prompts that get accurate, useful answers from AI.

A security attack where hidden instructions in user input hijack an AI's behaviour.

Compressing a model by using lower-precision numbers, so it runs faster and fits on smaller hardware.

A newer class of LLM that thinks internally before answering, trading speed for accuracy on hard problems.

A technique where the AI looks things up in your documents before answering.

Reinforcement Learning from Human Feedback — the technique that turned raw LLMs into helpful assistants.

Search that finds results by meaning, not just matching keywords.

AI that converts spoken audio into written text.

Sending the AI's answer word-by-word as it's generated, instead of waiting for the whole thing.

Data generated by AI to train or test other AI systems.

A hidden instruction that sets the AI's role, rules, and personality for the whole conversation.

A setting from 0-2 that controls how creative or random the AI's output is.

AI that reads written text aloud in a natural-sounding voice.

The unit of text an AI reads and generates — roughly ¾ of a word.

Letting an AI trigger external tools or APIs — like calling a function in code.

A sampling parameter that limits the AI to picking from the most likely next words.

The expensive process of teaching a model by adjusting its weights on huge amounts of data.

The neural network architecture behind almost every modern LLM.

A database designed to store and search embeddings by meaning.

Building software by describing what you want in plain English and letting AI write the code.

Chaining tools and AI together to complete an end-to-end business task without manual work.

Asking the AI to do a task with no examples — just an instruction.