The words you need to know, and the ones you can ignore
The jargon that matters, translated. And the jargon you can safely skip.
AI has a vocabulary problem. Most of the terms you'll encounter are either jargon designed to sound impressive, or marketing words with no fixed meaning. Here are a few terms worth understanding, and a short list of the ones you can safely skip.
The ones worth knowing
Large language model (LLM): The type of AI behind ChatGPT, Claude, and Gemini. A system trained on huge amounts of text that generates more text by predicting what comes next. When people say "AI" in a business context today, they usually mean this.
Prompt: The instruction you give the AI. "Write me a Slack message" is a prompt. So is a 500-word brief with examples. The quality of your prompt strongly affects the quality of the output.
Output: What the AI gives you back. Also called a response or completion. Whether it's good or bad depends on the prompt, the tool, and how lucky you got.
Hallucination: When the AI confidently produces something that isn't true. Made-up sources, invented quotes, fabricated facts. The polite term for "the AI lied with confidence." Worth knowing because it will happen to you.
Context window: How much information the AI can hold in mind at once during a conversation. Longer context windows mean you can paste in longer documents and have longer conversations before it starts forgetting earlier parts.
Token: The unit of text the AI works in. Roughly three-quarters of a word. Pricing and limits are often measured in tokens, not words. You don't need to count them, just know the term when it comes up.
Knowledge cutoff: The date after which the AI doesn't reliably know what happened. If you ask about events after the cutoff, you'll either get refusals, outdated information, or hallucinations. Always check the cutoff if you're asking about recent events.
Model: A specific version of an AI system. ChatGPT, Claude and Gemini each run on several models, and the names and version numbers change every few months. Different models from the same company can behave very differently.
Agent: An AI system that can take actions, not just generate text. It might browse the web, run code, or use other tools. Agents are early and unreliable as of 2026. Treat claims about them with scepticism.
API: The technical way one piece of software talks to another. When you hear about "building on top of Claude" or "integrating ChatGPT into our tool," they're using the API. You don't need to use one to use AI well.
Worth knowing
Large language model, prompt, output
Hallucination, context window, token
Training data, knowledge cutoff, model, system prompt
Safe to ignore
Transformer architecture
Parameters, and billions of parameters
Embeddings, vectors, vector databases
AGI and the singularity
Don't let vocabulary decide whether you're allowed in the room.
The ones you can mostly ignore
Some terms exist primarily to make people sound technical. You can engage with AI productively without ever using these words:
Transformer architecture: The underlying technical structure of LLMs. Interesting if you're a researcher. Useless if you're trying to write a client email.
Parameters / billions of parameters: A measure of model size. Companies advertise this number, but bigger doesn't reliably mean better. Ignore the numbers, judge the outputs.
Embeddings, vectors, vector databases: Useful concepts for engineers building AI products. Not useful for using AI tools.
Chain-of-thought: A specific prompting technique where you ask the AI to "think step by step." Useful in some contexts, oversold by most people who use the term.
AGI (artificial general intelligence): A hypothetical future AI that can do anything a human can. Lots of debate, no agreed definition, no clear timeline. Not relevant to your work right now.
Singularity: A speculative idea about AI becoming so advanced it changes society irreversibly. Strictly philosophical territory. Skip.
What to take from this
The point isn't to memorise the list. It's to stop letting vocabulary decide whether you're allowed in the room. If you know what a prompt is, what a hallucination looks like, and where a model's knowledge stops, you can use AI well. Everything else is detail you can pick up at the moment you actually need it.
Three things to do this week
01
Learn the handful that actually change what you do
Prompt, output, hallucination, context window, training cut-off. Five terms carry most of the weight.
02
Judge the output, not the spec sheet
Parameter counts and model names tell you nothing about whether it did your job well.
03
Ask why it wasn't said plainly
When someone leans on jargon to explain something simple, that's the question worth asking.
Conclusion
01
You need about fifteen terms to use AI well. The rest is jargon, marketing or speculation.
02
Heavy jargon is a signal about the writing, not about you.
03
The people who genuinely understand AI can almost always explain it in plain language.
Go to next