Guides
AI Terms Explained in Plain English: A Beginner's Glossary
20 essential AI terms explained in plain English: LLM, prompt, hallucination, token, RAG, agents, and more — one clear sentence each, no jargon.
AI articles, podcasts, and product pages are full of jargon: LLM, RAG, hallucination, token, agentic. Nobody hands you a dictionary, so every new term feels like a locked door. This guide unlocks the twenty you will actually run into, each in one plain sentence, followed by just enough context to make it stick.
Bookmark this page and come back whenever a new word pops up. Once you know these twenty, the rest of the AI conversation starts making sense.
The basics: what AI actually is
Artificial intelligence (AI)
Software that does tasks we usually associate with human judgment, like understanding language, recognizing images, or making predictions. Most AI in daily use is narrow AI: it does one kind of job well and nothing else. A spam filter is AI, but it cannot write your report.
Machine learning (ML)
A way of building software that learns patterns from examples instead of following hand-written rules. Show a model thousands of past orders and it learns which ones tend to be returned; the quality of the result depends on the quality of the examples.
Generative AI
AI that creates new content — text, images, audio, or video — rather than just sorting or predicting. ChatGPT, image generators, and AI music tools are all generative AI.
Large language model (LLM)
A model trained on huge amounts of text that learned to predict likely next words, which turns out to be enough to write, summarize, translate, and reason in language. ChatGPT, Claude, and Gemini are all powered by LLMs.
Chatbot
A program you talk to in plain language, usually powered by an LLM. The chat window is the interface; the model is the engine underneath.
How you talk to it
Prompt
The instruction or question you give an AI tool. Better prompts — specific, with context and a desired format — reliably produce better answers.
Prompt engineering
The practice of designing good prompts: giving background, examples, constraints, and a clear job description instead of a vague one-liner.
Zero-shot and few-shot
Zero-shot means asking the model to do a task with no examples. Few-shot means showing it a few input-and-output examples first, which usually improves accuracy on format-heavy jobs like sorting or labeling.
Chain-of-thought
Asking the model to work through a problem step by step before giving the final answer. It improves accuracy on math, logic, and multi-step reasoning — and it lets you spot where the reasoning went wrong.
System prompt
Hidden instructions given to the model before your conversation starts, setting its role, tone, and rules. App developers use them; when you tell a chatbot "act as my tutor," you are effectively writing your own.
How it works underneath
Token
The small unit of text a model reads and counts — a word, part of a word, or a character. Roughly, one token is about three-quarters of an English word, so 1,000 tokens is around 750 words. Both what you type and what the AI replies count as tokens, and many tools price by them.
Context window
How much text the model can hold in mind at once, measured in tokens. Once you exceed it, the earliest parts of the conversation start dropping out. A bigger context window means longer documents and longer chats stay coherent.
Parameters
The adjustable numbers inside a neural network, learned during training. Model names like 7B or 70B refer to billions of parameters. More usually means more capable, but also more expensive to run.
Training data
The text, images, and other material a model learned from. A model only knows what was in its training data (plus whatever it can look up), which is why it can be confidently out of date on recent events.
Fine-tuning
Extra training on a smaller, specialized set of examples to adapt a general model to one job — your company tone, a medical specialty, a coding style. It customizes behavior; it does not give the model new facts it can reliably quote.
Where it goes wrong — and how it gets fixed
Hallucination
A confident answer that is wrong or invented. The model predicts likely-sounding text; it has no built-in sense of truth. This is why dates, citations, and numbers from AI always deserve a check against a real source.
Retrieval-augmented generation (RAG)
Looking up relevant documents first, then answering based on what was found. Instead of relying only on training data, the model grounds its answer in your files, a help center, or the web — which reduces hallucinations and keeps answers current.
Grounding
Connecting an AI answer to real, checkable sources. Tools that cite their sources or search the web before answering are grounding their output; tools that answer from memory alone are not.
Guardrails
Rules layered around a model to block harmful, off-topic, or unsafe outputs — content filters, refusal rules, and human-review triggers. They are why the same question can get a refusal in one app and an answer in another.
The newer words you will keep hearing
AI agent / agentic AI
AI that does not just answer, but plans and takes steps toward a goal: searching, using tools, checking results, and adjusting. A chatbot answers your question; an agent books the flight, or at least tries to.
Multimodal AI
A model that works with more than one kind of input — text plus images, audio, or video. Upload a photo and ask what is wrong with it, or speak instead of typing, and you are using multimodal AI.
Model / frontier model
A specific trained system (like GPT-5 or Claude Sonnet) as opposed to the app wrapped around it. "Frontier model" means one of the most capable models available at the time — the label moves as new models ship.
Open-weight model
A model whose trained weights are published so anyone can download and run it, as opposed to a closed model you can only reach through the maker's app or API. Open-weight models can be run privately, but you still need hardware and know-how.
How the terms fit together
Here is one sentence that uses most of them at once: you write a prompt to a chatbot, which sends your words — counted in tokens — plus anything inside its context window to an LLM; if the app uses RAG, it first looks up your documents to ground the answer, and if it is agentic, it may take further steps on its own. When the answer sounds right but is not, that is a hallucination — the single most important word on this page, because it tells you when to double-check.
The bottom line
You do not need to memorize this glossary. You need to recognize the words when they appear and know which ones change what you should do: hallucination means verify, token means watch your usage, RAG means the answer is grounded in real documents, and agent means the tool may act, not just advise. With those instincts, AI jargon stops being a barrier and starts being useful shorthand.
Frequently asked questions
What is the difference between AI and an LLM?
AI is the broad field; an LLM is one kind of AI, specialized in language. Every LLM is AI, but not every AI is an LLM — a spam filter or a chess engine is AI without being a language model.
Why does AI hallucinate?
Because it predicts likely text rather than checking facts. Fluency is not accuracy, so confident-sounding answers still need verification whenever the facts matter.
What is a token in simple terms?
A chunk of text the model counts. Roughly four characters or three-quarters of a word, and the unit many AI tools use for usage limits and pricing.
Do I need to learn all these terms to use AI?
No. Learn prompt, hallucination, token, and context window first — those four change how you use the tools every day. The rest you can look up here when they appear.