Guides / AI basics

Guide · Basics

What Actually Is ChatGPT? AI Basics in Plain English

01First, the names

ChatGPT is a product, not the technology. The same way Hoover is one brand of vacuum cleaner, ChatGPT (from OpenAI) is one brand of large language model — usually shortened to LLM. Claude (Anthropic), Gemini (Google) and Copilot (Microsoft, built on OpenAI's models) are the other names you'll meet. They differ in personality and strengths, but under the bonnet they're the same kind of machine, and everything in this guide applies to all of them.

02How it works — no maths, promise

At its core, an LLM does one thing: given some text, it predicts what should come next.

Your phone's keyboard does a feeble version of this — type "see you" and it offers "later". An LLM is that idea taken to an absurd extreme. During training, the model was shown a vast slice of human writing — books, articles, websites, code — and gradually tuned, over months of computation, to get better and better at predicting the next word in all of it.

Here's the surprising part: to get really good at predicting the next word, the model had to absorb the patterns inside the text — grammar, facts, styles of argument, how a polite email differs from a legal letter, how code differs from poetry. Predicting "what a knowledgeable person would write next" ends up looking, in practice, very much like knowledge.

So when you ask ChatGPT to draft a quote follow-up email, it isn't looking one up. It's generating, word by word, the most plausible continuation of "a professional, friendly follow-up email about a quote" — drawing on the patterns of millions of emails it learned from. That's why the result feels written rather than retrieved.

03Why it makes things up

The same mechanism explains the technology's most important flaw. The model produces plausible text — and most of the time, plausible and true point the same way. But when the model doesn't know something, it doesn't stop and say so; the machinery keeps predicting, and out comes something that sounds right. A confident date. A convincing statistic. A legal case that doesn't exist.

This is called hallucination, and it isn't lying — there's no intent. It's the prediction machine doing exactly what it always does, just without the facts to anchor it. Two practical consequences for your business:

• Confidence is not evidence. The model sounds equally sure when it's right and when it's wrong. Anything factual that matters — figures, dates, regulations, citations — gets checked by a human.
• It's strongest where you supply the facts. "Summarise this contract" (the facts are in the contract) is far safer ground than "what does UK employment law say about X" (the facts are in its memory, which may be patchy or stale).

04Training cut-offs and web search

A model's built-in knowledge stops at its training cut-off — the date its training text was gathered. Ask about anything after that date and the honest answer is "it doesn't know", though, as we've seen, it may answer anyway.

Most of the big tools now work around this by searching the web mid-conversation and reading what they find. That's a genuine fix for recency, with one caveat: now the answer is only as good as the pages it happened to read. For anything that matters, ask the tool to show its sources — and click them.

05Why how you ask matters so much

The single biggest upgrade most people can make costs nothing: better prompts (a prompt is just what you type in). Remember, the model is predicting a continuation of what you give it. Give it a vague start, get a generic continuation. Give it a rich start, get a tailored one.

The mental model we teach: treat it like a bright new starter on their first morning. Capable, well-read, eager — and knowing absolutely nothing about your business, your customers or your standards until you tell them. You wouldn't say "write something for the customer" to a new starter; you'd brief them.

Same tool, same cost, dramatically different output. Role, audience, tone, constraints, and an example if you have one — that's most of prompt-writing in one sentence.

06A few terms decoded

Enough vocabulary to read the menu (for the full vocabulary, see The Plain-English AI Glossary):

• Model — the trained prediction machine itself. Tools usually offer a faster/cheaper one and a slower/smarter one.
• Prompt — whatever you type in. The briefing.
• Context window — the model's working memory for your conversation. Very long chats can overflow it, which is why a tool may "forget" something from much earlier — start a fresh chat for a fresh task.
• Tokens — the word-fragments the model actually reads and writes. Mostly invisible to you, but it's how usage gets measured and billed.
• Hallucination — see section 03. The reason humans stay in the loop.

07What this means in practice

Pulling it together: these tools are pattern engines, not oracles. That makes them superb at shaping work — drafting, rewording, summarising, explaining, transforming what you give them — and unreliable as a sole source of facts. Used that way, with a human checking what matters, they're one of the best productivity bargains available to a small business right now.

If you want the business case and where to start, that's our guide How Can AI Help My Small Business? And before pasting anything sensitive into one of these tools, read Is AI Safe to Use With My Business Data?