How to read this artificial intelligence dictionary
The NIST glossary warns that the meaning of many AI terms depends on context. This page helps you recognise those terms in the news and separate the technology from marketing labels. When a short definition is not enough, the links open a complete explainer.
Artificial intelligence (AI): broad label for systems that perform tasks associated with human abilities, such as recognising images, generating text or choosing between options. Generative AI: creates text, images, audio, video or code from patterns learned during training. AGI: disputed idea of an AI that performs well across a very broad range of tasks. No universal test can certify that a company has achieved AGI.
Machine learning: methods that learn relationships from data. Deep learning: type of machine learning that uses neural networks with many layers. Neural network: mathematical structure loosely inspired by connections in the brain. It is not a digital brain.
Model: trained mathematical system that turns an input into an output. LLM: large language model that predicts and generates sequences of text. ChatGPT is a product that can use several different models. Our guide to how ChatGPT works explains the distinction.
Training: process of adjusting a model's internal values with data and computation. Pretraining: broad initial learning. Fine-tuning: later adjustment with more specific examples. Post-training: techniques that shape behaviour, usefulness and safety after base training.
Prompts, tokens and context windows
Inference: stage in which a trained model processes a request and produces an answer. Parameter: internal value learned during training. The collection of those values is commonly called the model's weights. More parameters alone do not guarantee a better product. Checkpoint: saved copy of the weights at a particular point in training.
Prompt: all the information and instructions a model receives. It can include text, images, files, system rules and tool results, not just the sentence you type. Prompt engineering: designing that input for more reliable results through clear instructions, context and examples.
Token: unit into which a model divides information. It can be a word, part of a word or punctuation, as the OpenAI token guide explains. Our explainer covers what AI tokens are and why they matter. Context window: amount of information a model can consider in one request. If a long chat starts dropping instructions, read why AI forgets.
Long context: capacity for very large inputs. Context caching: reusing repeated input to reduce cost or time. Output tokens: units generated in the answer.
Embeddings, RAG and knowledge bases
Embedding: list of numbers that represents text, an image or other content so that similar items sit close together. Vector database: system designed to store those representations and search them by similarity.
RAG: retrieval-augmented generation. Before answering, the system finds relevant passages in a set of documents and adds them to the context. RAG does not retrain the model. Google describes the process as a way to support an answer with external information. Grounding: connecting the answer to specific sources or data to reduce unsupported claims.
Knowledge cutoff: approximate end date for some of a model's training data. It is not the newest information a product can use because web search, files and RAG can supply later material.
Agents, tools and automation
AI agent: system that receives a goal, chooses steps and uses tools with some autonomy. A chatbot that produces one answer is not automatically an agent. OpenAI's practical guide to agents separates the model, tools and instructions that make up the system. Our explainer shows how AI agents work.
Tool use or function calling: ability to request a structured search, calculation or action from another tool. It does not mean complete autonomy. Workflow: defined sequence that can include AI without letting it choose the route. Orchestrator: component that distributes work among models, tools or processes.
Human in the loop: a person reviews or authorises sensitive points. Autonomy: freedom to act without intervention. More autonomy does not automatically mean greater intelligence or safety.
Text, images, audio and video
Multimodal: model or system that handles several kinds of information, such as text and images. Vision: ability to interpret images or video. Speech-to-text: conversion of speech into text. Text-to-speech: generation of speech from text. Native audio: audio processed or generated inside the model rather than through several separate systems.
Diffusion model: family of generative models widely used for images and video. It learns to create content by gradually reversing noise. Latent space: compressed representation in which a model organises features. Upscaling: increase in image resolution. It can reconstruct detail, but cannot recover information that never existed.
Deepfake: synthetic or manipulated content that imitates a person's appearance or voice. Our guide to spotting AI text and images explains why no detector can provide certainty.
Reasoning, hallucinations and evaluations
Reasoning model: model optimised to spend more computation on complex problems before answering. Chain of thought: internal or exposed reasoning steps. A convincing explanation does not prove that the answer is correct. Test-time compute: computation used during inference to search, verify or spend longer on an answer.
Hallucination: false or unsupported output presented as valid. It does not involve an intention to lie. Confabulation: term sometimes used as an alternative. Calibration: how well expressed confidence matches actual accuracy. Uncertainty: explicit acknowledgement of limits or several possibilities.
Benchmark: test set comparing systems. Eval: evaluation for one capability or risk. Leaderboard: public ranking. A two-point gain matters only when the test represents the real task and avoids contamination.
Open source: software distributed under an open licence. Open weights: model whose weights are available even if its data, code or licence remain restricted. The terms are not synonyms. Proprietary or closed model: model whose provider controls access, usually through a product or API.
Guardrails: controls limiting outputs or actions. Jailbreak: attempt to bypass them. Prompt injection: malicious instructions hidden in processed content. Red teaming: adversarial testing to uncover failures.
Copilot: assistant that works alongside a person. It is a product name in some cases and a metaphor in others. Frontier model: label for systems near the current capability ceiling, without one shared boundary. State of the art: best known result on a particular test, not the best AI for every job. Choose by task with our guide to the best free AI tools.
The conversation starts here
Sign in with a supporter account to comment. Sign in
Nobody has commented yet. Want to go first?