Skip to content

Get the monthly AI briefing: what matters for your business, in plain English.Subscribe

Menu

Plain-English AI glossary

AI Made Clear

Plain-English definitions of the AI terms you will come across, from A to Z.

Start with the basics

A

Acceptable use
The specific list of what is fine, what needs sign-off, and what is off limits.
AI agent
An AI given a goal, some tools, and permission to take steps on its own.
AI literacy
The skills and knowledge needed to use AI systems appropriately and understand their risks and limitations.
AI register
A record of the AI systems an organisation uses, with owners, use cases, data and risk ratings.
AI use policy
A document setting out how employees may and may not use AI tools at work.
Audit trail
A record of what was asked, what came back, and who approved it.

B

Benchmark
A standard test used to compare one model against another.
Bias
Skew carried in from the training data that then shows up in the output.

C

Chain of thought
Asking the model to work through its reasoning before it answers.
Context window
How much text a model can hold in mind at once, your instructions included.
Copilot (AI assistant)
AI that sits alongside a person and suggests, rather than deciding.

D

Data protection impact assessment (DPIA)
A formal privacy check carried out before launch.
Data residency
Which country your data is actually stored and processed in.

E

Embeddings
Text turned into numbers, so things can be found by meaning rather than keyword.
EU AI Act
EU regulation establishing risk-based rules for the development and use of AI systems.
Explainability
Being able to show how and why a decision was reached.

F

Few-shot prompting
Including a few worked examples so the model copies the pattern you want.
Fine-tuning
Training a general model further on your own data so it suits your work.
Foundation model
A large general-purpose model that other products are built on top of.

G

Generative AI
AI that produces new content rather than just sorting or scoring what exists.
Guardrails
Limits built around a tool governing what it may do or say.

H

Hallucination
When a model states something false with complete confidence.
Human in the loop
A person reviews or approves the output before anything actually happens.

I

Inference
The moment a trained model actually runs and produces your answer.
ISO/IEC 42001
An international standard specifying requirements for an AI management system.

L

Large language model
A model trained on an enormous amount of text to predict what words come next.

M

Model card
A vendor's own summary of what a model is for, and what it is not.
Model drift
Quality quietly falling away as the world moves on from the training data.
Multimodal
A model that handles more than text: images, audio, documents, video.

O

Orchestration
Joining several models, tools and steps into one dependable process.

P

Parameters (weights)
The numbers inside a model that hold everything it learned.
Prompt
The instruction you give the model. What you type is the prompt.
Prompt engineering
Writing and refining prompts until the results are dependable.

R

Retrieval-augmented generation (RAG)
The model looks things up in your documents before it answers.

S

Shadow AI
The use of AI tools for work without the knowledge or approval of the organisation.
Synthetic data
Generated data used to stand in for real records.
System prompt
Standing instructions set behind the scenes that shape every answer.

T

Temperature
A setting that decides how predictable or inventive the answers are.
Token
The chunks of text a model reads and writes, each roughly three quarters of a word.
Training data
The material a model learned from. It shapes everything the model produces.
Training opt-out
Switching off a vendor's right to learn from what your people put in.

U

Unstructured data
Everything not in tidy rows and fields: emails, documents, calls, notes.

V

Vector database
Where embeddings are stored so the right document is found quickly.