AI glossary: 30 terms from the news, explained

Basics: what AI is

• Generative AI: AI that creates new things like text, images, music or video. ChatGPT, Claude and Gemini are examples.
• LLM (large language model): the program inside a chatbot that writes text, built by learning from a huge amount of writing.
• Chatbot: a program you talk to by sending messages back and forth. Most AI chatbots today run on an LLM.
• Multimodal: able to understand or create more than text, such as photos, audio and video. Uploading a photo and asking "what breed is this cat?" is an example.
• Prompt: the instruction or question you give an AI. The same task can give very different results depending on how you write the prompt.
• Token: the smallest piece of text an AI processes. It sets chat length limits and API pricing.
• Context window: how much text an AI can consider at once. It works like short-term memory and is counted in tokens.
• Algorithm: a fixed set of steps for getting something done. In AI news it often means the method a program uses to recommend or decide something.

Models: what happens inside AI

• Model: the trained AI itself. "GPT model" and "Claude model" each come in named versions.
• Training data: the text, images and other material an AI learned from. It is at the center of many copyright disputes.
• Parameters: the huge number of values inside a model that are adjusted during training. People describe model size as "hundreds of billions of parameters," but more is not always better.
• Pretraining: the first stage of training, where the model reads a huge amount of text and learns to predict what comes next. A pretrained model alone does not follow instructions well, so more training follows.
• Fine-tuning: training an already-trained model on extra material so it fits a field or a style.
• Reasoning model: a model that works through a problem before answering, which helps with complex questions. The tradeoff is slower replies and possibly more usage.
• Open-source (open-weight) model: a model whose files are published so you can download and run or modify it. How open it is, and the terms of use, differ by model, so check the license.
• GPU: a graphics processing unit. It is good at doing many calculations at once, which is why it is used to train and run AI. It is also why chips come up so often in AI news.

Usage: how people use AI

• RAG (retrieval-augmented generation): having the AI look up relevant documents first and answer based on them. It is used to bring in up-to-date information or a company's internal files.
• Agent: an AI that takes a goal and carries out several steps on its own, like searching, organizing and creating files.
• MCP: a common standard for connecting AI to other apps such as Google Drive, Notion and calendars.
• API: a way for programs to use each other's features. Developers use an AI company's API to put AI in their own apps, usually paying per token.
• Vibe coding: building software by describing what you want to an AI in plain language instead of writing the code yourself.
• System prompt: the base instructions an app's maker gives the AI in advance. It is separate from what you type.
• Temperature: a setting for how random the answers are. Higher tends to give more varied, creative answers. Lower tends to give more consistent ones.
• Subscription / usage limits: paying a flat monthly fee but with a cap on how much you can use per hour or per day.

Safety: what to watch for

• Hallucination: when an AI states something false as fact. Check numbers and sources yourself.
• Deepfake: a video or voice made with AI to look or sound like a real person doing or saying something they never did. It is used in scams and misinformation.
• Watermark: a visible or invisible mark showing that something was made by AI. Having one does not catch every fake.
• Prompt injection: an attack that hides instructions in a web page or document to trick an AI into doing something unintended.
• Bias: when skewed training data makes an AI's answers lean in a particular direction.
• AGI (artificial general intelligence): a hypothetical AI that can do nearly any intellectual task a person can. Companies define it differently, so check the definition behind any claim that it has arrived. There is no agreement that today's chatbots are AGI.

Checking that a definition is accurate

This field moves fast, and companies sometimes use the same term in slightly different ways. Before quoting a term in something important, check an official glossary too. Anthropic's published glossary covers terms like context window, fine-tuning, LLM, MCP, RAG and tokens with short definitions.

Related terms

When you do not know a term

If a news story uses an AI term you do not know, ask an AI "explain this in one sentence a middle schooler could follow." Check any numbers or dates separately.

FAQ

Keep going

• Claude Code: MCP for developers
• Things to make with AI