A large language model: AI trained on vast text to predict and generate human-like language. ChatGPT, Claude, and Gemini are all LLMs.

What is a hallucination in AI?

When an AI generates confident-sounding but false information. LLMs predict plausible text, not verified facts — so always check important claims.

Home / Guides · September 2026

AI Jargon Glossary

Every AI term you'll meet as a beginner, explained in one or two plain sentences. Bookmark this — you'll be back.

Term	Plain meaning
AI / artificial intelligence	Software that performs tasks needing human-like judgment: understanding language, recognizing images, making predictions.
Generative AI	AI that creates new content — text, images, code, audio — rather than just classifying or predicting.
LLM	Large language model: AI trained on vast text to understand and generate human-like language. ChatGPT, Claude, Gemini.
Prompt	The instruction you give an AI. Better prompts = better results; it's a skill, and it's learnable.
Token	A chunk of text (roughly ¾ of a word) — the unit AI reads and bills by. Long chats consume thousands.
Context window	How much text the AI can consider at once. Bigger window = longer documents and conversations remembered.
Hallucination	Confident-sounding false output. LLMs predict plausible text, not verified facts — check important claims.
Training data / cutoff	The text the model learned from, frozen at a date. Anything newer is unknown unless the tool browses the web.

Term	Plain meaning
System prompt	Hidden instructions shaping the AI's behavior (tone, rules). You can often add your own custom instructions.
Temperature	A creativity dial: low = predictable and factual, high = surprising and varied. Most apps hide this.
Few-shot prompting	Showing the AI 2–3 examples of what you want before asking. Massively improves format compliance.
Chain of thought	Asking the AI to reason step by step ("think through this carefully") — improves complex answers.
RAG	Retrieval-augmented generation: the AI looks up your documents before answering. How chatbots "know" your files.
Agent	AI that takes multi-step actions (browse, click, run code) toward a goal, not just one reply. Powerful, needs supervision.
Fine-tuning	Extra training on specific data to specialize a model. Overkill for beginners; prompting usually suffices.
Embeddings	Numbers capturing text meaning, enabling semantic search ("find similar"). Infrastructure you'll never touch directly.

Term	Plain meaning
Foundation model	A big general-purpose model others build on (GPT, Claude, Gemini, Llama). The engine under many apps.
Open weights	Models you can download and run yourself (like Llama). "Open" varies — check the actual license.
Multimodal	Handles multiple input types: text plus images, audio, or video. Most flagships are now multimodal.
Alignment / guardrails	Training and filters steering AI away from harmful outputs. Why models refuse some requests.
Jailbreak	Tricking a model past its guardrails. A security research topic — and a terms-of-service violation on most platforms.
Benchmark	Standardized tests scoring models (bar exams, coding tasks). Useful directionally; gameable and narrow.
API	The paid pipe developers use to embed AI in apps. You pay per token; playgrounds let you test without coding.
Copilot	AI assistant embedded in a tool (coding, Office, browser). The generic pattern; also Microsoft's brand name.
Diffusion model	Image-generation AI that refines noise into pictures (Midjourney, DALL-E, Firefly). Different tech from LLMs.

A large language model — AI trained on vast text to understand and generate human-like language.

Confident-sounding false output. Always verify important AI claims against real sources.