The short answer
If you do not write code, nearly everything useful about AI sits behind a chat box. You type a request in ordinary English, you read what comes back, and you check it before you rely on it. That is the whole interface. The jargon that keeps stopping you (tokens, models, agents, API keys) exists because most AI writing is aimed at engineers, not because the subject demands it, and the parts that matter to you take about an afternoon.
Why does every AI guide sound like it was written for someone else?
Because it was. The first people to write about these tools were the people building them, and they wrote for their colleagues: parameter counts, benchmark tables, notes on which library to install. All accurate, and none of it any use to a practice manager, a head of year, a plumber or a regional sales lead. Then came the people selling courses, who had noticed that jargon reads as expertise and that a reader who feels out of their depth is easier to sell to.
So if you have read four articles and finished each one feeling less capable than when you started, that is a writing problem rather than a you problem. There is no prerequisite you missed, and no maths module you skipped at school that would have made this click.
A quick test for any AI article: if it tells you to get an API key in the first three paragraphs, it was not written for you. Close the tab. Nothing you need this month sits behind a key.
What do the words actually mean?
Five terms block more people than all the others combined. Here they are with the fog removed.
In plain English
- Prompt:
- Whatever you type into the box: a question, an instruction, a pasted document, or all three. No special syntax, no magic words. A clear request, written as you would brief a competent temp, is a good prompt.
- Model:
- The particular AI you are talking to, the way a car has a model. ChatGPT, Claude and Gemini come from different companies and have different personalities. All of them work identically from your side of the screen.
- Hallucination:
- When the tool states something untrue in a completely confident voice. Not lying, not broken: it is producing text that fits the shape of a good answer, and sometimes that answer is wrong.
- Token:
- A chunk of text, roughly three quarters of a word. Companies count tokens for billing and length limits. In a free chat box you can forget this word exists.
- Agent:
- An AI allowed to do things rather than only reply: search the web, open a file, book something. Worth recognising when a product advertises it, safe to leave alone for now.
That is the vocabulary. Almost everything else you meet this year is a variation on those five.
What are the only three things you need to understand?
Three facts, and between them they explain nearly every strange thing these tools do.
One: it predicts plausible text, it does not look things up
It is not consulting a database of facts and reporting back. It is producing the words that most plausibly follow your request, based on an enormous amount of text it has read. Usually the most plausible words are also the true ones, which is why it works at all. Occasionally they are not, and you get a beautifully formatted citation for a report that does not exist. So it is superb at shaping things you already know, and unreliable at supplying facts you do not.
Two: it does not know you unless you tell it
Every conversation starts from nothing. It does not know your job, your team, your last email or what you asked it on Tuesday. Nearly every disappointing answer traces back to this: the request was fine, the context was missing.
I am a [your role] at a [type and size of organisation] in the UK.
I am writing for [who will read this] and they care most about [what].
Our house style is [formal / plain / warm / brief].
Here is the task: [what you want]
Here is my raw material: [paste your notes, figures or draft]
Ask me any questions you need before you start.
Those first three lines are the difference between generic filler and something you can actually send.
Three: it sounds the same when it is wrong
There is no wobble in the voice. A well-sourced summary and an invented statistic arrive in the same tidy prose, with the same air of quiet competence. People hedge when they are unsure. This does not, so you cannot use tone as a truth detector.
What can you do with it today, in your actual job?
Office and admin work. Turn the messy notes from a fifty minute meeting into a summary with decisions, owners and dates. Then reverse it: paste a long email thread and ask what you are actually being asked to do. See AI Productivity at Work.
Teaching. Three versions of the same worksheet at different reading levels, then a mark scheme, then a note home explaining what the class covers this half term. See AI for Educators.
Managing a team. Rehearse the difficult conversation before you have it: ask the tool to play a defensive team member and push back on your wording, then rewrite your opening line. See AI for Managers.
Running a small business. Turn a scribbled job into a proper quote email, or draft the polite third chase for an unpaid invoice. See AI for Small Business.
Job hunting. Rewrite a vague CV bullet into an evidence-based one using facts you supply, then tailor it to a specific advert. Never let it invent the facts. See AI-Powered Job Search.
Sales and support. Draft the twelfth version of a reply you have written eleven times, in your own tone, with this particular customer's detail in it.
What can you safely ignore?
This is the most useful section here, so take it as permission rather than advice.
Version numbers. Whether you are on the newest release changes very little for any of the above. Use whatever the free tier gives you.
Benchmark scores. Leaderboards measure exam-style tasks. They do not tell you which tool writes a better message to a disappointed customer. Try two, keep the one you prefer.
The is it really intelligent debate. Genuinely interesting, and entirely irrelevant to whether it can tidy your meeting notes.
Prompt engineering as a mystique. There are no secret phrases. Anyone selling a pack of one thousand magic prompts is selling clear instructions with a price tag attached.
Anything needing an API key, a terminal or an install. That is developer territory: real, useful, and not required for almost all of the benefit.
If a piece of AI advice makes you feel behind, check whether it is telling you to do anything at all. Most of the anxious content is about the industry, not about your work.
What is the one habit that matters more than any technique?
Check the output before you rely on it, and scale the checking to the cost of being wrong.
A first draft of an internal email needs no verification, because you are the check. A figure in a board pack, a claim in a customer contract, a legal or tax or medical detail, a name you are about to spell in public: those you confirm at the source. Not by asking the AI whether it is sure, which only produces more confident text, but by looking it up somewhere that can be held responsible.
List every factual claim, number, name, date and quotation in what
you just wrote.
For each one, say where it came from: my material, general knowledge,
or an assumption you made. Mark anything I should verify myself
before sending this to [audience].
Do not rewrite anything. Just give me the list.
Two minutes, and it catches most of what would otherwise embarrass you. The free AI Safety and Verification course is built entirely around this habit.
Where should you start?
Open a free chat tool and give it a real task from today, not a test question. Something you were going to do anyway and are slightly dreading. Paste your actual material, tell it who you are, and judge the result the way you would judge a first draft from a keen new colleague: useful, imperfect, needs your eye on it.
If you want a path rather than a browser tab full of half-read articles, we built one. Start at the beginner's route, take Getting Started with ChatGPT as your first hour, and browse the full course list when you know what you want next. Everything is free and nothing needs an account.