Most Commonly Used AI Terms & What They Mean

Six terms worth actually understanding if you’re working anywhere near AI or LLMs right now:

1. Tokens

The smallest units of text a model processes. A prompt gets broken into tokens before the model can work with it — roughly, 1 token ≈ 4 characters, or about ¾ of a word in English. A single word can sometimes span two tokens.

2. Context window

The model’s working memory — how much text (measured in tokens) it can hold and reference at once. A bigger context window means it can consider more of a conversation or document at a time.

3. Hallucinations

When a model generates incorrect or fabricated information and presents it as fact — usually the result of gaps or biases in training data. This isn’t theoretical: in 2025, Deloitte’s Australian arm had to issue a partial refund on a report after it included AI-generated citations to research that didn’t exist, along with a fabricated court quote.

4. Retrieval-Augmented Generation (RAG)

A method for pulling information from a private knowledge base — a company’s internal documents, a database — rather than relying only on what the model learned during training. This is how a support chatbot can correctly answer “how much leave do I have left” by retrieving your actual leave record alongside the policy document.

5. Grounding

The broader practice of tying a model’s output to verifiable, factual sources rather than letting it generate from memory alone — reducing the chance of invented content. This is often what’s happening when you see a model run a live web search mid-conversation.

6. Model Context Protocol (MCP)

Introduced by Anthropic in late 2024, MCP is a standardized way for LLMs to connect to external data, apps, and services — often described as “a USB-C port for AI applications.” It’s what lets an AI model move from static knowledge to actually taking action across different tools.

Originally published on Medium.