What a model actually is

The one-sentence version

A modern AI chatbot is a large language model, a system that, given some text, predicts what text is most likely to come next. That's the whole trick. It reads your message, then produces its reply one chunk at a time, each chunk chosen because it's a plausible continuation of everything that came before.

Those chunks are called tokens: roughly, pieces of words. "Understanding" might be one token, or two ("under" + "standing"). The model doesn't think in sentences; it thinks in "what token comes next, and next, and next", stringing them together faster than you can read.

That sounds almost too simple to be useful. It isn't. Predicting the next word well turns out to require a rough working model of grammar, tone, facts, code, and the shape of a good argument. But it's still prediction. Hold on to that. It explains everything else in this lesson.

Training versus using

There are two very separate moments in a model's life, and beginners often blur them.

Training happened once, in the past, before you ever opened the tool. The model was shown an enormous amount of text and adjusted itself, over and over, to get better at predicting the next token. This produced a huge set of internal numbers, its "weights". When you hear a model has a training cut-off (say, some point in 2024 or 2025), that's the last time its underlying knowledge was baked in.

Using it, what you do in the chat box, doesn't teach it anything new. Each time you send a message, the model runs its prediction over your text and its own instructions, produces a reply, and then, in the most basic sense, forgets. It isn't looking anything up in a database as it answers. It's generating a likely response from patterns learned during training.

This is why a model can confidently tell you about a person or product but get a recent detail wrong: its "knowledge" is a frozen snapshot, blurred into patterns, not a live encyclopedia.

Why it isn't a search engine (or a database)

A search engine has an index of real pages and hands you links to them. A database stores exact records and returns them unchanged. A language model does neither by default. It has no filing cabinet of verified facts to open. When it gives you a date, a statistic or a quote, it is producing the most likely-looking date, statistic or quote, which is very often right, and sometimes confidently, fluently wrong. When a model invents something that isn't true, we call it a hallucination.

Two consequences follow, and they matter for the rest of the course:

• Fluency is not accuracy. The tool is optimised to sound right. A smooth, well-structured, authoritative answer tells you nothing about whether it's true.
• It can't reliably tell you how sure it is. Asking "are you certain?" often just produces another confident paragraph. Confidence is generated text too.

When it can look things up

The important modern wrinkle: most 2026 tools can now search the web during an answer when you turn that on (or when they decide to). With web search on, the tool fetches real pages, reads them, and writes its reply grounded in what it found, usually with links. That doesn't make hallucination impossible, but it moves the tool much closer to "reads real sources, then summarises" and away from "guesses from memory". You'll learn the details tool-by-tool later; for now, just know the switch exists and changes the game.

A worked example: the same question, two ways

Imagine you're at Fernway and someone asks you a factual, time-sensitive question.

What is the current standard rate of UK VAT, and when did it last change?

With web search off, a typical model answers instantly and fluently: it names a rate and a date. For a stable, well-known fact this is usually correct, but you have no source, and if the figure had changed after its training cut-off, it would have no way to know. It would still sound completely sure.

Now the same question with search on:

Search the web and tell me the current standard rate of UK VAT and when it last changed. Give me the GOV.UK link you used so I can verify it.

Now the tool fetches a page, quotes the figure, and hands you a link. The answer might be identical, but now it's checkable. That's the difference in a nutshell: same clever guesser, but one version is anchored to reality and one is floating free.

A good follow-up, once you have a grounded answer, is to make it useful:

Using that VAT figure, write two plain-English sentences I could put in a customer email explaining how VAT affects their invoice total.

Try it now

Common mistakes

• Treating it as a search engine. Asking a bare model for a fresh statistic and pasting the result into a report. Without search on and a source checked, you're quoting a guess.
• Trusting the confident tone. The single biggest beginner error is reading fluency as accuracy. A model has no "I'm unsure" wobble in its voice; it sounds equally certain when it's right and when it's inventing. Over-trusting a smooth answer is how wrong figures end up in real documents.
• Asking "are you sure?" as a check. That just generates more text. Real verification means opening a source, not interrogating the model.
• Assuming it knows today's news. Its baseline knowledge stops at its training cut-off. Anything recent needs web search on.

Keeping current

Model names, training cut-offs and how aggressively each tool searches the web change constantly. When a claim about "what the model knows" matters, check the tool's own model documentation, for example OpenAI's model release notes, rather than trusting a number you memorised. Treat every specific fact in this lesson as true on 13 July 2026 and worth re-checking after that.