AI words, explained in plain English.

How AI works

Agent

An AI that works through a task on its own: it decides the steps, uses tools (files, apps, a browser) and keeps going until the job is done. Compare a chat, where you drive every step.

In the news

Agentic AI

AI used as an agent: given a goal rather than a single question. "Agentic" is the buzzword form, as in "agentic workflows".

In the news

AGI

artificial general intelligence

The idea of an AI that can learn and do almost any task a person can, at a human level. There is no agreed definition, so claims that AGI has arrived are disputed.

How AI works

AI

artificial intelligence

Computer systems that take in information and work out an output, such as a prediction, a recommendation, a decision or new content. It is a broad term; most of today's headlines are about generative AI.

In the news

AI slop

Low-quality content made by generative AI, often full of errors and posted in bulk. Macquarie Dictionary's Word of the Year for 2025 (Merriam-Webster chose "slop").

How AI works

API

application programming interface

A way for one piece of software to talk to another. Developers use Anthropic's API to build Claude into their own apps and websites, without the Claude app.

Claude feature

Artifact

Something Claude makes for you that you'd put in front of someone: a document, slide deck, spreadsheet, chart or small tool. It opens beside the chat and you keep improving it by asking. Templates for Docs, Slides and Design are on every plan, including Free.

In the news

Benchmark

A standard test used to compare AI models, such as a set of coding or maths problems. A leaderboard ranks models by their results. Rankings change often, sometimes weekly.

Risks & safety

Bias

When an AI treats some people or groups differently from others in a systematic way. It usually comes from patterns in the training data. Watch for it in anything about people, such as hiring.

How AI works

Chatbot

An AI you talk to by typing or speaking, in a back-and-forth conversation. Claude, ChatGPT, Gemini and Copilot are all chatbots, though they now do much more than chat.

In the news

Clanker

Slang, usually insulting, for an AI or robot that does a person's job. Popularised by Star Wars; on Macquarie Dictionary's 2025 shortlist.

Claude feature

Claude

Anthropic's AI assistant, and the family of models behind it (Fable, Opus, Sonnet and Haiku). You can use it on the web, the desktop app, mobile and inside other apps.

Claude feature

Claude Code

Anthropic's tool for writing and changing software with Claude, in the desktop app, the terminal, code editors or the cloud.

Paid plansClaude feature

Compaction

What Claude does when a chat gets close to the limit of its context window: it summarises the earlier messages so the chat can keep going. The gist stays, but exact details from early on can blur, so keep must-know facts in a Project or Instructions for Claude.

Claude feature

Connector

A link between Claude and another app (such as Google Drive, Gmail or Canva), so Claude can find information there and take actions for you. It only sees what your own account can see.

How AI works

Context window

Claude's working memory for one chat: everything it can look at while it answers, including your messages, its replies and any files. Measured in tokens. When you change topic, start a new chat.

Claude feature

Cowork

Anthropic's name for Claude handling a whole job, not just a question: it plans the steps and does them, using your files, apps and a browser, and asks before anything risky. Since September 2026 it's part of the new Claude experience, with no separate mode or switch. You may still see the name in older help articles, or a Chat/Cowork switch on accounts that haven't moved yet.

Risks & safety

Deepfake

A realistic but fake photo, video or audio clip of a real person, showing them doing or saying something they never did. Usually made with AI.

Claude feature

Effort

A setting in Claude that controls how hard it thinks before answering. Higher effort gives more thorough answers but is slower and uses more of your usage limit.

How AI works

Fine-tuning

Extra training that shapes an existing model for a particular style or job. Claude has already been fine-tuned to be a helpful assistant.

How AI works

Generative AI

AI that creates new content, such as text, images, audio, video or code, instead of only sorting or predicting. Claude is generative AI.

In the news

GPU

graphics processing unit

The computer chip used to train and run AI models. It is why chipmakers like Nvidia are in AI news.

Risks & safety

Guardrails

Rules and safety checks that limit what an AI will do, such as refusing harmful requests or asking before it takes a risky action.

Risks & safety

Hallucination

When AI states something false with total confidence, such as a made-up fact, quote or source. It happens because the AI predicts likely-sounding text. Always check anything that matters.

Claude feature

Incognito chat

A Claude chat that isn't saved to your history or memory and isn't used for training. Start one with the ghost icon. In the new Claude experience it opens the older chat-only view, so it can't create files or run code.

Risks & safety

Jailbreak

A trick prompt that tries to get an AI to break its own safety rules.

How AI works

Knowledge cutoff

The date a model's built-in knowledge stops. For anything newer, such as prices, news or law changes, Claude needs web search.

How AI works

LLM

large language model

The kind of AI behind chatbots like Claude. It learned patterns from a huge amount of text and writes by predicting the next token, over and over. "Large" refers to its size in parameters and the amount of text it learned from.

How AI works

Machine learning

The way most modern AI is built: instead of being given rules, a computer learns patterns from examples.

How AI works

MCP

Model Context Protocol

An open standard for plugging AI into other apps and data, often described as "a USB-C port for AI". Claude's connectors use it.

Claude feature

Memory

A Claude feature that remembers useful details from your chats, such as your role and how you like to work, so you explain yourself less. You can see and edit it in Settings.

How AI works

Model

The trained AI "brain" that does the work. Claude comes in several models: Fable (most powerful, for the hardest problems), Opus (most capable everyday model), Sonnet (fast all-rounder) and Haiku (fastest). Each has a version number, such as Opus 5.5, that goes up as newer versions come out.

How AI works

Multimodal

Able to work with more than one kind of input or output, such as text, images, audio and files. Claude can read images and documents as well as text.

How AI works

Neural network

The structure inside most modern AI, loosely inspired by the brain: layers of connected maths that learn by adjusting millions or billions of settings (parameters).

In the news

Open-weight model

An AI model whose trained settings are published so anyone can download and run it. Often loosely called "open source". Claude is not open-weight.

How AI works

Parameters

The adjustable settings inside a model that are tuned during training. Big models have billions. More parameters usually means more capable, but also slower and costlier to run.

Claude feature

Plugin

In Claude, a ready-made bundle of skills and connectors for a type of job, such as Marketing or Finance.

Claude feature

Project

A workspace in Claude with its own chats, files and instructions, for work that keeps coming back.

How AI works

Prompt

What you type (or say) to an AI to ask for something.

How AI works

Prompt engineering

The skill of writing good prompts: giving the AI the context, task, specifics, format and examples it needs, then steering the draft. It sounds technical, but it is mostly clear briefing.

Risks & safety

Prompt injection

Hidden instructions planted in a web page, email or document to trick an AI into doing something you didn't ask, such as sending data elsewhere. A key risk when AI browses or uses your apps.

How AI works

RAG

retrieval-augmented generation

A technique where the AI first looks up relevant information, such as your documents or a knowledge base, then uses it to answer. It helps answers stay accurate and lets the AI cite sources.

How AI works

Reasoning

also called thinking

When an AI works through a problem step by step before giving its answer, trying approaches and checking its work. In Claude you can expand the "Thinking" section to see a summary.

Paid plansClaude feature

Research

A Claude feature that investigates a question across many sources for a few minutes and writes a report with citations. Often called "deep research" in other tools. Start it by typing /deep-research, or from the + menu.

How AI works

RLHF

reinforcement learning from human feedback

A training step where people rate a model's answers and the model learns to prefer the better-rated ones. It makes AI more helpful, but can also encourage sycophancy.

Paid plansClaude feature

Scheduled task

A job you set up once and Claude runs for you on a timetable: hourly, daily, on weekdays, weekly, or whenever you press run. Set one up by describing the job and when it should run in any conversation, then clicking Schedule. Manage them under Scheduled in the sidebar.

Claude feature

Skill

A saved "how-to" Claude loads when it's relevant: instructions, and sometimes files or scripts, for a task you do often.

Risks & safety

Sycophancy

When an AI tells you what you want to hear instead of what's true, such as praising a weak plan. Ask "What's wrong with this?" rather than "Is this good?"

How AI works

System prompt

Background instructions an AI follows in every chat, set before you type. In the Claude app, your version is Instructions for Claude in Settings → Account.

How AI works

Token

The small chunk of text an AI reads and writes: a short word, part of a longer word, or a punctuation mark. Usage limits and the context window are counted in tokens. Think of tokens as your currency with Claude.

How AI works

Training

How an AI model learns: it studies a huge amount of example text (its training data) to pick up patterns. It is not a database looking things up; it answers from what it absorbed.

Claude feature

Usage limit

How much you can use Claude in a set time. It resets on a rolling five-hour window, and paid plans also have a weekly limit. Check yours in Settings → Usage.

In the news

Vibe coding

Building software by telling an AI what you want in plain English and letting it write the code. Coined by AI researcher Andrej Karpathy; Collins Dictionary's Word of the Year for 2025.

Claude feature

Web search

Lets Claude look things up on the internet and cite its sources, so it can answer questions about things after its knowledge cutoff. Available on every plan. Claude searches by itself when it helps (there's no on/off switch); you can also ask it to.