AI fundamentals: what it is, and what it is not

The kind of AI you meet at work

"AI" is a broad word, but at work it almost always means one thing right now: generative AI, and in particular large language models, the technology behind ChatGPT, Microsoft Copilot, Claude, and the assistants built into your other tools. These are systems that generate text, and sometimes images or code, in response to a request written in plain language.

A handful of words come up again and again. Here is all the vocabulary you actually need.

Model

The trained system that produces the answer. Think of it as the engine.

Prompt

What you type in. Your question, instruction, or the text you paste.

Token

A chunk of text the model reads and writes in, roughly a short word or part of one.

Training data

The huge body of text the model learned patterns from, up to a fixed cut-off date.

How a language model actually answers

This is the single most useful idea in the whole course. A language model does not look up an answer in a database. It predicts likely text, one piece at a time, based on patterns it saw during training. Given your prompt, it works out what words tend to come next, then the next, and the next, until it has a full response.

This is why AI is brilliant at some things and unreliable at others. Predicting fluent, well-structured language is exactly what it was built to do, so drafting, summarising, rephrasing, and explaining come naturally. Recalling a precise fact, a figure, or a source is a different job, and the model will happily produce something that reads correct whether or not it is correct.

Why it sounds so confident, even when it is wrong

A model has no built-in sense of true or false, and no feeling of doubt. It produces the most plausible-sounding text every time, in the same steady, confident voice, whether the answer is solid or invented. When it states something false in that confident tone, people call it a hallucination. It is not lying, and it is not broken. It is doing precisely what it does: predicting plausible text.

Two more limits worth holding onto:

• It has a cut-off. Unless a tool is connected to live data, the model only knows patterns up to its training date. Ask about last week and it may guess.
• It does not know your world. On its own it has never seen your files, your customers, or your policies. Connecting it to your systems is a separate step, and the subject of module five.

The AI you will actually run into

You do not need to track every product name, but it helps to recognise the shapes:

• Chat assistants: a box you type into, like ChatGPT or Claude. General purpose.
• Embedded copilots: AI built into a tool you already use, such as your email, documents, or code editor.
• AI agents: AI that does not just answer but takes actions in real systems, such as sending an email or updating a record. This is a meaningful step up in both usefulness and risk, and it is the whole of module five.
• Image and media generators: same idea as text, applied to pictures, audio, or video.

What AI is not

Clearing up four common misreadings saves a lot of trouble:

• It is not a search engine. It generates text, it does not retrieve verified pages, unless a tool explicitly adds that.
• It is not a database of facts. Do not treat a figure it gives you as sourced until you have checked.
• It is not up to date by default. Assume a knowledge cut-off unless told otherwise.
• It is not a person. It has no intent, no memory of you between sessions unless designed to, and no accountability. The accountability stays with the humans who use it.