The Beginner's AI Glossary: Every Term People Actually Use, Explained in Plain English

Every conversation about AI now comes pre-loaded with vocabulary that nobody stops to define. Your coworker says a chatbot "hallucinated," a headline promises "agentic AI" will reorganize your job, and a product page brags about a "1-million-token context window" as if that means something to a normal person. This guide is the fix: the essential AI terms explained in plain English — not all 200 of them, just the twenty-odd words people actually use in 2026.

A quick promise before we start: no math, no architecture diagrams, and no pretending that "transformer" is something you need to understand to use ChatGPT. The goal is that the next time someone drops "RAG" or "fine-tuning" into a meeting, you know exactly what they mean — and, more usefully, when they're using the word wrong.

The basics: AI, models, and LLMs

Artificial intelligence (AI) is the umbrella term for software that performs tasks we associate with human intelligence — understanding language, recognizing images, making decisions. It's deliberately vague, which is why marketers love it.

Machine learning (ML) is the dominant way we build AI today: instead of programming rules by hand, you feed a system enormous amounts of data and let it find patterns on its own. Nearly everything called "AI" in 2026 is machine learning underneath.

A model is the actual product of that learning — the trained system you interact with. GPT-5.5, Claude, and Gemini are all models. When someone says "the model got it wrong," they mean the specific AI system, not the company.

A large language model (LLM) is a model trained on massive amounts of text to predict what words come next. That sounds trivial, but predict-the-next-word at sufficient scale turns out to produce systems that can write code, summarize contracts, and argue about movies. Generative AI is the broader category: any model that creates new content — text, images, audio, video — rather than just classifying or ranking existing stuff.

Tokens, prompts, and context windows: AI terms explained for daily use

These three words come up constantly because they describe the actual mechanics of talking to a chatbot.

• Prompt — whatever you type into the AI. The instruction, the question, the pasted document. Prompt engineering is the craft of writing prompts that get better results, though in 2026 models are forgiving enough that it's less of a dark art than it used to be.
• Token — the chunks of text a model actually processes. A token is roughly three-quarters of an English word; "unbelievable" might be three tokens. Models are priced, limited, and benchmarked in tokens, which is why the word is everywhere.
• Context window — the model's working memory: the total amount of text (in tokens) it can consider at once, including your conversation history. When a chatbot "forgets" what you said twenty messages ago, you've scrolled past the edge of its context window.
• Inference — the act of actually running a trained model to produce an answer. Training happens once, at enormous cost; inference happens every time you hit enter.

Hallucination, grounding, and RAG: when AI gets it wrong

Hallucination is the industry's polite word for a model confidently making things up — fake citations, invented court cases, plausible-sounding statistics that don't exist. It happens because LLMs generate statistically likely text, not verified facts. As TechCrunch's AI glossary puts it, hallucinations stem from gaps in training data — the model fills the hole with something fluent instead of admitting ignorance.

The one thing to remember about every AI tool you use: an LLM doesn't retrieve facts, it predicts plausible text. Hallucination isn't a bug that will be patched next quarter — it's a side effect of the core design, which is why you verify anything that matters.

Grounding is the family of techniques for tethering a model's answers to real sources. The most common one is RAG (retrieval-augmented generation): before answering, the system searches a database or the web, pulls in relevant documents, and generates its answer from those. It's why modern chatbots can cite sources and answer questions about yesterday's news. Perplexity built an entire product around this idea.

Agents and agentic AI: the words of 2026

An AI agent is a system that doesn't just answer a question but completes a multi-step task on its own — researching a topic, writing and running code, filling out forms, booking things. The key difference from a chatbot: an agent loops. It plans, acts, checks its work, and tries again without you prompting each step.

Agentic AI is the adjective form and the buzzword of the moment. It describes systems that operate with limited human supervision across multiple tools. The hype is real but so is the deployment: Gartner expects 40 percent of enterprise applications to include task-specific AI agents by the end of 2026, up from under 5 percent a year earlier.

Two related terms you'll hear alongside it. Tool calling (or function calling) is the mechanism that lets a model use external software — a calculator, a search engine, your calendar. And MCP (Model Context Protocol) is an open standard, originally from Anthropic and since adopted by OpenAI, Google, and Microsoft, that gives agents a common way to plug into those tools.

Under the hood: training, parameters, and fine-tuning

Training is the months-long, wildly expensive process of feeding a model data so it learns patterns. Parameters (or weights) are the billions of internal numbers that get adjusted during training — when you hear a model has "70 billion parameters," that's a rough proxy for its size and capability, not a quality guarantee.

Fine-tuning means taking an already-trained model and training it a bit more on specialized data — legal documents, medical records, your company's support tickets — so it performs better in that niche. RLHF (reinforcement learning from human feedback) is a specific refinement step where humans rate model outputs and the model learns to prefer the well-rated ones. It's a big part of why chatbots feel helpful rather than feral.

The party-conversation tier: AGI, multimodal, and open weights

AGI (artificial general intelligence) is the hypothetical system that matches or exceeds humans at most economically valuable work. Nobody agrees on the definition, which makes it perfect for both fundraising decks and doom threads. Treat any confident AGI timeline with suspicion.

Multimodal means a model handles more than text — it can see images, hear audio, or generate video. Most flagship models in 2026 are multimodal by default. Reasoning models are variants that "think" step by step before answering, trading speed and cost for better performance on hard problems; they're why some chatbot responses now come with a visible pause.

Open weights (often loosely called open source) means a model's parameters are published so anyone can download and run it, unlike closed models such as GPT-5.5, which you can only access through OpenAI's paid tiers.

FAQ

What's the difference between AI, machine learning, and an LLM?

They're nested categories. AI is the broad goal of making software do intelligent things; machine learning is the technique of learning patterns from data rather than hand-coded rules; and an LLM is one specific kind of machine-learning model, trained on text. Every LLM is machine learning, and all of it counts as AI — but plenty of AI (like image classifiers) isn't an LLM.

Why do AI models hallucinate, and can it be fixed?

LLMs generate the most statistically plausible next words, so when they lack real knowledge they produce something fluent instead of stopping. Grounding techniques like RAG, citations, and better training have reduced the problem substantially, but no current model is hallucination-free. The practical fix is on your end: verify names, numbers, and citations before you rely on them.

Is an AI agent just a chatbot with a different name?

No. A chatbot responds to one message at a time and waits for you; an agent takes a goal and works through multiple steps on its own — searching, using tools, checking results, and correcting course. The distinction matters because agents can act on the real world (sending emails, editing files, spending money), which raises the stakes when they get something wrong.