AI glossary

43 AI terms explained in plain English, from agents and tokens to RAG, MCP and open weights. Where we’ve written a full explainer, the term links to it.

A

AGI (Artificial general intelligence): A hypothetical AI system able to match or exceed people across most intellectual tasks, rather than excelling at specific ones. There is no agreed test for it, so claims that a system is or is not AGI depend on the definition being used.
AI agent: A system in which an AI model decides its own next steps—calling tools, reading the results and acting again—until it reaches a goal or stops. A chatbot answers a question; an agent can carry out a multi-step task such as researching, booking or fixing code. Read our full explainer →
Alignment: The work of making AI systems pursue the goals their developers and users intend, and avoid harmful behaviour. It covers training methods, testing and rules about what a model should refuse to do.
API (Application programming interface): A way for software to talk to other software. Most AI models are offered to businesses through an API: a program sends text, images or files and receives the model’s output, usually paying per token.

B

Benchmark: A standard test used to compare AI models, such as a set of coding tasks or exam questions. Scores depend heavily on how the test is run, and results published by a model’s own developer are not independent verification. Read our full explainer →

C

C2PA and Content Credentials: An open standard for attaching a signed record of how a piece of media was made and edited. A Content Credential shows provenance—who or what created a file and how it changed—but it does not prove the content is true. Read our full explainer →
Chain of thought: Intermediate reasoning steps a model writes out before giving an answer. Encouraging step-by-step reasoning often improves accuracy on maths and logic problems; “reasoning models” are trained to do this by default.
Compute: The processing power used to train and run AI models, usually measured in GPU-hours or in the number of specialised chips available. Access to compute is a major constraint for AI developers and a focus of national programmes such as the IndiaAI Mission. Read our full explainer →
Context window: The maximum amount of text, measured in tokens, that a model can read and write in a single request. Anything outside the window is invisible to the model unless something puts it back in. Read our full explainer →

D

Data centre: A building that houses servers, storage and networking equipment. AI data centres need large amounts of electricity and cooling, which is why power supply increasingly shapes where AI capacity is built. Read our full explainer →
Deepfake (Synthetic media): Audio, images or video generated or altered by AI to look or sound like real people or events. Several countries, including India, now have rules requiring platforms to label or act on certain synthetic media. Read our full explainer →
Diffusion model: The type of model behind most AI image and video generators. It learns to turn random noise into an image step by step, guided by a text description.
Distillation: Training a smaller “student” model to imitate a larger “teacher” model’s outputs, so the small model keeps much of the capability at lower cost. Some AI providers forbid using their outputs to distil competing models.

E

Embedding: A list of numbers that represents the meaning of a piece of text, an image or other data, so that similar items have similar numbers. Embeddings power semantic search and retrieval-augmented generation.
EU AI Act: The European Union’s law regulating AI systems by risk level, with separate obligations for prohibited practices, high-risk systems, transparency and general-purpose AI models. Its duties apply on different dates rather than all at once.

F

Fine-tuning: Further training an existing model on a smaller, specialised dataset so it performs better at a particular task or follows a particular style. It changes the model’s weights, unlike prompting or retrieval.
Foundation model: A large model trained on broad data that can be adapted to many tasks, such as a large language model. Products are often built on top of a foundation model from another company.

G

GPU (Graphics processing unit): A chip originally designed for graphics that can perform many calculations in parallel, which makes it well suited to training and running AI models. Counts of GPUs are often used as a shorthand for AI capacity.
Guardrails: Rules and filters around an AI system that block or redirect unwanted inputs and outputs—for example refusing certain requests, or checking an agent’s actions before they run.

H

Hallucination: When an AI model states something false or made up with apparent confidence, such as an invented citation or fact. Grounding answers in retrieved sources and asking the model to quote them reduces, but does not eliminate, hallucinations.

I

Inference: Running a trained model to produce an output, as opposed to training it. Most of the ongoing cost of an AI product is inference: every question answered uses compute.

L

Large language model (LLM): An AI model trained on very large amounts of text to predict and generate language. LLMs power chat assistants, coding tools and many agents.

M

MCP (Model Context Protocol): An open standard that lets AI applications connect to external tools and data—files, databases, business apps—through a common interface, so each service needs one connector rather than one per AI product.
Mixture of experts (MoE): A model design in which only some of the model’s parameters (“experts”) are used for each token. It lets a model have many total parameters while using fewer “active” ones per step, though all the weights still need to be stored. Read our full explainer →
Model card: A document published with an AI model describing what it is for, how it was trained and evaluated, and its known limitations. Model cards are written by the developer, so their claims are self-reported.
Multimodal: Able to work with more than one type of data—for example text, images, audio and video—in the same model.

O

Open weights: A model whose trained parameters can be downloaded and run by anyone. Open weights are not necessarily “open source”: the training data and code may be private, and the licence may restrict use.

P

Parameters (Weights): The numbers inside a model that are adjusted during training and determine its behaviour. Parameter count is a rough measure of model size, not of quality. Read our full explainer →
Prompt: The instructions and input given to an AI model. Clear, specific prompts with examples usually produce better results.
Prompt injection: An attack in which instructions hidden in content an AI reads—a web page, email or document—trick it into doing something the user didn’t ask for. It is one of the main security risks for AI agents.

Q

Quantisation: Storing a model’s numbers at lower precision, such as 4-bit instead of 16-bit, to make it smaller and faster to run on ordinary hardware, usually with a small loss in quality.

R

RAG (Retrieval-augmented generation): A technique that searches a collection of documents for passages relevant to a question and adds them to the prompt, so the model answers from those sources. It is the standard way to make AI answer questions about private or recent information. Read our full explainer →
Reasoning model: A model trained to work through problems step by step before answering, spending more computation on harder questions. It tends to do better on maths, coding and planning, at higher cost and latency.
Red teaming: Deliberately trying to make an AI system misbehave—produce harmful content, leak data or take unsafe actions—so the problems can be fixed before or after release.
RLHF (Reinforcement learning from human feedback): A training method in which people rate a model’s responses and the model is adjusted to prefer the higher-rated ones. It is widely used to make chat assistants more helpful and to follow safety rules.

S

Small language model (SLM): A language model small enough to run on a laptop, phone or modest server, trading some capability for lower cost, speed and privacy.

T

Token: The unit of text a model reads and writes—often part of a word. In English one token is roughly three-quarters of a word; many other languages, including Indian languages, often need more tokens for the same text. AI usage is usually priced per token.
TPU and NPU: Specialised AI chips. A TPU (tensor processing unit) is Google’s accelerator for data centres; an NPU (neural processing unit) is a smaller chip in phones and laptops for running AI on the device.
Training: The process of building a model by adjusting its parameters on large amounts of data. Training a frontier model takes vast compute; running the finished model is called inference.
Transformer: The neural-network architecture behind most modern language models, introduced in 2017. Its “attention” mechanism lets the model weigh how every part of the input relates to every other part.

V

Vector database: A database designed to store embeddings and quickly find the items most similar in meaning to a query. It is a common building block of RAG systems.

W

Watermarking: Embedding a hidden, detectable signal in AI-generated text, images or audio at the moment it is created. It only works when the AI provider adds it, and heavy editing can weaken it.

Z

Zero-shot and few-shot: Asking a model to do a task with no examples (zero-shot) or with a handful of examples in the prompt (few-shot). A few good examples often improve results without any retraining.

Spotted a definition that could be clearer? Tell us.