AI glossary
61 AI terms explained in one or two plain sentences.
A
- Agent
- An AI system that can plan steps and use tools (search, code, apps) to complete a task with some autonomy.
- Algorithm
- A set of step-by-step rules a computer follows to solve a problem.
- Alignment
- Work to make AI systems behave according to human intentions and values.
- API
- Application Programming Interface — a way for software to send requests to an AI model and receive answers.
- Artificial intelligence (AI)
- Computer systems performing tasks that normally need human intelligence, such as understanding language or recognising images.
- Attention
- The transformer mechanism that lets a model weigh how relevant each part of the input is to every other part.
B
- Backpropagation
- The method that works out how much each weight contributed to an error so it can be adjusted during training.
- Benchmark
- A standard test used to compare AI models' performance.
- Bias
- Systematic unfairness in AI outputs, often inherited from training data or design choices.
C
- Chain-of-thought
- Prompting or training a model to work through intermediate reasoning steps before answering.
- Chatbot
- Software you interact with through conversation; modern chatbots are often powered by LLMs.
- Chunking
- Splitting documents into smaller passages so they can be searched and retrieved.
- Classification
- Assigning an input to a category, such as spam or not spam.
- Computer vision
- AI that interprets images and video.
- Context window
- The maximum amount of text (in tokens) a model can consider at once.
D
- Dataset
- A collection of examples used to train or test a model.
- Deep learning
- Machine learning using neural networks with many layers.
- Diffusion model
- A type of generative model, common for images, that learns to turn noise into a picture step by step.
E
- Embedding
- A list of numbers representing the meaning of text or images so similar items sit close together.
- Evaluation (evals)
- Systematic testing of AI outputs against a test set and scoring rubric.
F
- Few-shot prompting
- Including a few examples in a prompt to show the desired output.
- Fine-tuning
- Further training a pre-trained model on specific examples to specialise its behaviour.
- Foundation model
- A large model trained on broad data that can be adapted to many tasks.
G
- Generative AI
- AI that creates new content such as text, images, audio or code.
- GPU
- Graphics processing unit — hardware widely used to train and run AI models.
- Gradient descent
- An optimisation method that repeatedly adjusts weights to reduce error.
- Grounding
- Tying a model's answer to specific trusted sources.
- Guardrails
- Rules, filters or checks that keep AI behaviour within acceptable limits.
H
- Hallucination
- A fluent, confident AI output that is false or unsupported.
- Human in the loop
- A design where a person reviews or approves AI outputs before they take effect.
I
- Inference
- Running a trained model to produce an output.
L
- Large language model (LLM)
- A neural network trained on large amounts of text to predict and generate language.
- Latency
- How long a model takes to respond.
M
- Machine learning
- Systems that learn patterns from data rather than following only hand-written rules.
- Model
- The trained mathematical system that turns inputs into outputs.
- Multimodal
- Able to work with more than one kind of data, such as text, images and audio.
N
- Natural language processing (NLP)
- The field of AI concerned with understanding and generating human language.
- Neural network
- A model made of layers of connected units whose weights are learned from data.
O
- Open-weight model
- A model whose trained weights are published so others can run or adapt it.
- Overfitting
- When a model memorises training data and performs poorly on new data.
P
- Parameter
- An internal value (weight) learned during training; large models have billions.
- Pre-training
- The initial large-scale training of a model on broad data.
- Prompt
- The instruction and context you give an AI model.
- Prompt chaining
- Breaking a task into a sequence of prompts, each using the previous output.
- Prompt engineering
- The practice of designing prompts that produce reliable, useful outputs.
- Prompt injection
- Hidden or malicious instructions in content that try to hijack an AI system's behaviour.
R
- Reasoning model
- A model trained to spend more computation working through problems step by step before answering.
- Reinforcement learning
- Learning by trial and error using rewards.
- Retrieval-augmented generation (RAG)
- Retrieving relevant documents and giving them to a model so it answers from those sources.
- RLHF
- Reinforcement learning from human feedback — using human preference ratings to shape model behaviour.
S
- Semantic search
- Search based on meaning rather than exact keywords, often using embeddings.
- Supervised learning
- Training on examples paired with correct answers.
- System prompt
- Background instructions that set an AI assistant's behaviour for a whole conversation.
T
- Temperature
- A setting controlling how random or varied a model's outputs are.
- Token
- A chunk of text — a word, part of a word or punctuation — that models read and write.
- Training data
- The examples a model learns from.
- Transformer
- The neural-network architecture behind most modern language models, built around attention.
U
- Unsupervised learning
- Finding patterns in data without labelled answers.
V
- Vector database
- A database designed to store embeddings and search them by similarity.
W
- Weights
- The learned numbers inside a neural network that determine its behaviour.
Z
- Zero-shot prompting
- Asking a model to do a task without giving examples.