Every field defends itself with vocabulary, and AI defends itself better than most. Here are the terms you actually need, defined in plain English, with the honest translation where the official one is evasive. Pin it to the noticeboard; nothing here requires a computer science degree, because none of the underlying ideas is that complicated.
AI (artificial intelligence). An umbrella term so broad it's nearly useless — it has covered everything from chess computers to spam filters. When people say it today, they almost always mean the next entry.
Generative AI. Software that produces new material — text, images, music, video — rather than just sorting or finding existing material. The kind that writes your minutes, and the kind at the centre of the fight over artists' work. Both, confusingly, under one name.
LLM (large language model). The engine inside ChatGPT and Claude. A system trained on colossal amounts of text until it becomes very good at predicting what word plausibly comes next. Everything miraculous and everything unreliable about these tools follows from that one sentence.
Model. One specific trained system — the actual "brain", as opposed to the app you type into. Companies ship new ones constantly, with names like software updates. You don't need to track them, merely to know that "the new model" explains why the thing suddenly behaves differently.
ChatGPT, Claude, Gemini, Copilot. Products — friendly shopfronts in front of models, from OpenAI, Anthropic, Google and Microsoft respectively. We've mapped the whole confusing family here.
Prompt. Whatever you type in. "Prompt engineering" is the grand term for phrasing requests clearly — a skill arts administrators, who spend their lives writing careful emails to difficult people, tend to possess in embarrassing abundance.
Hallucination. The industry's word for when AI states something false with total confidence — an invented statistic, a venue that doesn't exist, a quotation nobody said. Not lying (lying requires knowing better), just fluent guessing. The single most important term on this list: it's why a human checks everything.
Training. The process of building a model by feeding it data — for the big models, much of the internet, which is precisely where the consent problem lives. Related: training data, the material ingested; and the setting worth finding in your own account, which decides whether your chats become training data. (More on that here.)
Context window. The model's working memory — how much it can hold in mind at once. Why you can paste in a whole board pack and it copes, and why, in a very long conversation, it eventually forgets how the conversation started. Like all of us at the end of a festival.
Token. The unit models actually read — roughly a short word or piece of one. Matters to you only because it's how usage gets measured and billed. The syllable, monetised.
Agent. The current buzzword: AI that doesn't just answer but takes multi-step actions — searching, filling in, booking, doing. Treat with measured scepticism and never give one your card details.
Open source (of models). A model whose workings are published for anyone to download and run, rather than rented from a company. Matters for the sector mainly as a signal that this technology won't stay locked in four corporate buildings.
Algorithm. A recipe followed by a computer. The word predates AI by decades and mostly signals, when used interchangeably with it, that a speaker is improvising.
AGI (artificial general intelligence). The hypothetical future system as capable as a human at everything. Definitionally slippery, perpetually imminent, and — this is the useful part — not what's in your laptop. Every current tool is narrow: dazzling at reformatting, hopeless at taste.
GPT. Stands for something technical you will never need to know. Nobody at the conference knows it either. This is the noticeboard's final lesson: the vocabulary is a fence, not a foundation — step over it.
The image above was AI-generated — as ever, we'll always tell you.