LEARN · AI GLOSSARY

AI glossary: 100 terms in plain language

Every term has a one-line definition, a real example and a link to try the idea yourself — because you understand AI words best by using them.

AI basics 5

Artificial intelligence (AI): Computer systems that perform tasks we associate with human thinking, such as understanding language, recognising images or making predictions.
Deep learning: Machine learning using neural networks with many layers, which can learn complex patterns in text, images and sound.
Generative AI: AI that creates new content — text, images, audio, video or code — based on patterns learned from training data.
Machine learning: A way of building AI where the system learns patterns from examples instead of following hand-written rules.
Neural network: A model made of layers of simple connected units whose connection strengths (weights) are adjusted during training.

Models 9

Closed model: A model available only through the maker’s app or API; the weights are not published.
Foundation model: A large general-purpose model trained on broad data that can be adapted to many tasks.
Large language model (LLM): A very large neural network trained on huge amounts of text to predict the next word, which lets it write, summarise, translate and answer questions.
Model: The trained file of numbers (weights) that turns inputs into outputs. Different models have different strengths, sizes and costs.
Multimodal: A model that can work with more than one type of input or output, such as text plus images or audio.
Open-weight model: A model whose trained weights are published so anyone can download, run or fine-tune it, subject to its licence.
Parameters: The adjustable numbers inside a model. More parameters can mean more capability but also more cost and slower responses.
Reasoning model: A model trained to work through problems step by step before answering, often slower but better at maths, logic and planning.
Small language model: A compact model designed to run cheaply or on a device, trading some capability for speed and privacy.

Prompting 10

Chain of thought: Asking a model to reason step by step, which can improve answers on multi-step problems.
Few-shot prompting: Including a few worked examples in the prompt so the model copies the pattern.
Max tokens: A limit on how long the model’s reply can be, which also caps cost.
Prompt: The instructions and context you give an AI model. Clear prompts get more useful, checkable results.
Prompt engineering: The practice of designing prompts — role, task, context, format and checks — to get reliable results.
Prompt template: A reusable prompt with blanks to fill in, so good instructions can be used again and again.
Role prompting: Telling the model who to act as, so it adopts suitable knowledge and tone.
System prompt: Hidden instructions that set a model’s role and rules before the user’s message.
Temperature: A setting that controls randomness. Low values give focused, repeatable answers; high values give more varied ones.
Zero-shot prompting: Asking a model to do a task with no examples, relying on instructions alone.

How it works 22

API: An application programming interface: a defined way for programs to talk to each other, such as sending a prompt to a model.
API key: A secret code that identifies you to an API. Treat it like a password.
BYOK (bring your own key): Using your own API key with a tool, so you control the account, limits and costs.
Context window: The maximum amount of text (in tokens) a model can consider at once, including your prompt and its reply.
Embedding: A list of numbers that represents the meaning of text, so similar meanings are close together.
Endpoint: A specific URL an API listens on for a type of request.
Fine-tuning: Further training an existing model on a smaller, specific dataset to specialise it.
Function calling: A feature that lets a model request that your code run a defined function, such as looking up an order.
GPU: A graphics processing unit: a chip that does many calculations in parallel, used to train and run AI models.
Inference: Running a trained model to get an output. Every chat reply is an inference call.
JSON: A simple text format for structured data, used by almost every API.
Latency: The delay between sending a request and getting a response.
OpenAI-compatible API: An API that copies the request format popularised by OpenAI, so one client can talk to many providers.
Quantisation: Compressing a model’s numbers to use less memory, so it runs faster or on smaller hardware with a small quality cost.
Rate limit: A cap on how many requests you can make in a time window, such as requests per minute.
Retrieval-augmented generation (RAG): Fetching relevant documents first and giving them to the model, so answers are grounded in your sources.
Streaming: Sending a model’s reply piece by piece as it is generated, so users see text appear immediately.
Structured output: Asking a model to reply in a strict format such as JSON so code can read it reliably.
Token: A chunk of text (often part of a word) that models read and write. Pricing and limits are counted in tokens.
Training: The process of adjusting a model’s weights using data so it learns patterns.
Training data: The examples a model learns from. Its quality, coverage and bias shape what the model can do.
Vector database: A database that stores embeddings and quickly finds the most similar ones.

Safety & trust 16

Alignment: Efforts to make AI systems follow human intentions and values.
Benchmark: A standard test set used to compare models. Useful, but real tasks may differ.
Bias: Systematic unfairness in outputs, often reflecting gaps or stereotypes in training data.
Data privacy: Protecting personal information from misuse, including what you paste into AI tools.
Deepfake: AI-generated audio, video or images that convincingly imitate a real person.
Evaluation (evals): Systematic tests that measure how well a model or prompt performs on defined tasks.
Grounding: Tying a model’s answer to provided sources so claims can be traced and checked.
Guardrails: Rules and checks around a model that block unsafe inputs or outputs.
Hallucination: When a model states something false or invented as if it were true.
Human in the loop: Keeping a person responsible for reviewing or approving AI outputs before they take effect.
Jailbreak: Attempts to get a model to break its safety rules through clever prompts.
Model card: A document describing a model’s intended use, training data, limits and evaluation results.
Prompt injection: An attack where text inside content tricks a model into ignoring its instructions.
Red teaming: Deliberately trying to make a system fail or misbehave to find weaknesses before others do.
Redaction: Removing or masking sensitive details from text before sharing it.
Watermarking: Embedding hidden signals in AI-generated content so it can be identified later.

Data & ML 17

Accuracy: The share of predictions that are correct. Misleading when classes are imbalanced.
Baseline: A simple reference result that a model must beat to be worth using.
Classification: Predicting a category, such as spam or not spam.
Clustering: Grouping similar items without predefined labels.
Confusion matrix: A table showing correct and incorrect predictions for each class.
Data cleaning: Fixing or flagging errors, duplicates and missing values before analysis.
Data leakage: When information from the test set or the future sneaks into training, making results look better than they are.
Feature: An input variable a model uses to make predictions.
Label: The correct answer attached to a training example.
Notebook: An interactive document mixing code, results and notes, such as Jupyter or Colab.
Overfitting: When a model memorises training data and performs poorly on new data.
Pandas: A popular Python library for working with tables of data.
Precision and recall: Precision: how many flagged items were right. Recall: how many right items were flagged.
Regression: Predicting a number, such as a price or a quantity.
Synthetic data: Artificially generated data that mimics real data, used for testing or privacy.
Train/test split: Keeping some data aside, unseen during training, to measure real performance.
Uncertainty: How unsure a prediction is. Good systems report it instead of faking precision.

Apps & automation 11

AI agent: A system where a model plans steps and uses tools (search, code, APIs) to complete a goal with some autonomy.
Browser extension: A small add-on that changes how your browser works.
Chatbot: A program that holds a conversation, from simple FAQ bots to LLM assistants.
Copilot: An AI assistant built into a tool to help as you work, such as code or document suggestions.
Model Context Protocol (MCP): An open standard for connecting AI assistants to tools and data sources through one common interface.
No-code: Building apps or automations with visual tools instead of writing code.
Progressive web app (PWA): A website that can be installed like an app and work offline.
Serverless: Running code without managing servers; the platform runs it on demand.
Static site: A website made of fixed files (HTML, CSS, JS) that can be hosted free.
Webhook: A URL that receives automatic messages from another service when an event happens.
Workflow automation: Connecting apps so tasks run automatically when something happens.

Creative AI 10

Aspect ratio: The shape of an image, width to height.
Copyright and AI: Legal questions about training data and who owns AI-assisted work. Rules differ by country and are changing.
Diffusion model: An image model that learns to turn random noise into a picture step by step.
Hook: The first line or seconds of content that make people stop and pay attention.
Negative prompt: Things you tell an image model to avoid.
Speech-to-text: Turning spoken audio into written text (transcription).
Text-to-image: Generating images from a written description.
Text-to-speech (TTS): Turning written text into spoken audio.
Upscaling: Increasing an image’s resolution using AI to add detail.
Voice cloning: Creating a synthetic copy of a real person’s voice. Requires consent; often misused in scams.