What is AI? A plain explanation for people who have never used it

The short answer, and what people actually mean

AI is software that has read an enormous amount of human writing and learned the patterns in it well enough to produce new writing that fits. You type a request in ordinary language. It writes back: a draft, a summary, an explanation, a plan. That is the whole product. Everything else you have heard is either a detail sitting on top of that, or a story about something else entirely.

The term technically covers older things too, such as the spam filter in your inbox, the face unlock on your phone, and the system deciding which video plays next. Nobody means those in conversation. When a colleague says she used AI to fix a paragraph, or your nephew says he asked AI about his knee, they mean a chatbot: ChatGPT, Claude, Gemini, or one of their relatives. The product-level view of that category lives in what an AI chatbot is. This article sits one level above it, on the idea itself. If the vocabulary is the part that puts you off, model and prompt and token and context, the plain-language glossary is short and unpretentious.

Signing up for one takes an email address and about a minute, and every major app has a free tier, so nothing here needs a card. Before you read further, open any AI app and type this: "Explain what you are and how you work, in four sentences, to someone who has never used AI before." Leave the answer open in another tab. By section four you will have something concrete to test it against, which beats taking anyone's word for it.

How it learned to do that

The training process is dull to describe, and that is a good sign. Take a very large pile of text: books, websites, manuals, forum arguments, court filings, recipes. Hide a word. Make the system guess it. Tell it whether the guess was right. Repeat an absurd number of times. Nothing more exotic is happening. What comes out is a system with an extremely fine sense of what usually comes next in human writing.

So when you ask it who won the 1998 World Cup, it is not opening a reference book. It is producing the text that most plausibly follows your question, one small piece at a time. It happens to be right about France, because that fact appears in its training text thousands of times in a consistent form. It is right for the same reason it is fluent, not for some other, sturdier reason. Hold onto that sentence. Almost everything strange about how AI behaves falls out of it.

• It writes well immediately. Fluency is exactly what it was trained on, so a first draft arrives in decent shape without any effort from you.
• It sounds identical when it is wrong. Nothing inside it separates a fact seen ten thousand times from one it is assembling out of plausible-looking parts. The confidence is the same either way.
• It has a cutoff. Training stopped on some date. Ask about last Tuesday and, unless the app runs a live web search for you, you are asking a system that was not there.
• It is shaky at arithmetic. Long multiplication is a procedure, not a pattern in prose. Models handle it better now by writing out steps or calling a calculator, but the underlying instinct is still to produce a number that looks correct.
• Your phrasing steers the answer. Your words decide which patterns it reaches for, which is why the same question asked two ways comes back at two different levels of quality. Talking to AI is mostly that single skill.

What it is genuinely good at

The tasks it does best share a shape. The form of a good answer is already well established in the world, and you can tell within about ten seconds whether what came back is any good. Rewriting, summarizing, translating, explaining a document, and getting past a blank page: that is most of the writing in an ordinary week.

What you already have	What to ask for	Why this one works
An email you have rewritten four times	"Rewrite this so it is warm but still clearly says no"	Polite refusals are one of the most common patterns in written English
A letter from an insurer or a landlord	"Explain this in plain language, then list what I have to do"	Turning jargon into ordinary words is pattern work, and you can check it against the letter in front of you
A 60-page PDF you were never going to read	"Summarize the argument in eight bullets, then pull out every number"	It reads the whole file in seconds, and you can spot-check the bullets you care about
A trip you keep not planning	"Plan four days in Seville for two adults who like food and walking"	Thousands of itineraries sit in the training text, and an imperfect one costs you nothing
A message to a client in another language	"Translate this into Spanish, polite but not stiff"	Tone survives here in a way it did not with older translation tools

Look at what is missing from that table: not one row asks the machine to be the authority. You already hold the material, the judgment, or both. It takes the part you find tedious, the blank page, the second draft, the reformatting, the sixty pages, and leaves the deciding to you. Everyday tasks worth handing over has a longer list. One habit to build early: when the first answer misses, say what is wrong in one short sentence instead of starting over. "Too formal, and cut the last paragraph" fixes more than a rewritten request usually does.

What it is bad at, for exactly the same reason

Now the honest part, because you will meet this in your first week anyway. It will tell you something false in the same calm, well-organized voice it uses for everything else. Not rarely, and not only about obscure topics. It will invent a section number in a lease, a study with believable authors and a believable year, a product feature that does not exist, a quotation nobody ever said. There is no tell: no hedge in the wording, no wobble in the tone. The word for this is hallucination, and why AI gets things wrong walks through the failure modes one at a time, because recognizing which kind of wrong you are looking at is most of the defense.

Three further limits, all growing from the same root. Recency: unless the app is searching the web live, it does not know the past few months, and it will discuss them anyway using older material. Counting and arithmetic: it can reason its way through a calculation and still hand you a total that is off by a digit, so verify any number you would act on. Anything specific to you: your contract, your dose, your local rules, your deadline. It has read a million documents that resemble yours and not one of yours.

The rule worth keeping is this. Do not use it where being 95 percent right is worthless. Medication, legal deadlines, tax figures, the amount on an invoice, the spelling of a name in a speech you are giving. In those cases nearly right is not partial credit, it is simply wrong, and the fluent presentation makes the error harder to catch than a bad Google result would be. Everywhere else, 95 percent is a strong first draft that you then spend four minutes correcting. If your hesitation is less about accuracy and more about what happens to the things you type, that is a fair and separate question, and is AI safe to use treats it as one.

Your first five minutes

Reading about this has a low ceiling. Open an app and type one of the four below, adapted to something real from your own week. Real matters: a made-up test question gets you a made-up-feeling answer, and teaches you nothing about whether this is useful to you.

1. The awkward email. "Here is an email I need to send. Rewrite it so it is polite, about half as long, and still clearly says no." Paste your draft underneath.
2. The letter you dread. "Explain this letter in plain language, then tell me exactly what I have to do and by when." Paste or upload the thing from the insurer, the landlord, or the tax office.
3. The plan you keep postponing. "I have four days in Lisbon in October with one teenager and roughly 700 dollars for food and activities. Plan it day by day, then tell me what you assumed about us."
4. The question you feel silly asking. "Explain how a mortgage rate is actually set, as if I am intelligent but have never bought a house. Then give me the three questions to ask a lender."

A good answer does the task instead of describing how it would do the task. It holds the format you asked for. Where you left something out, it either asks you or names the assumption it made. If what comes back is a padded, over-headed, essay-shaped thing, that is usually a symptom of a vague request rather than a weak model, so add who it is for, how long you want it, and what you plan to do with it. Models also differ more than the marketing suggests, in voice and in how carefully they follow a long instruction, so the fastest way to find your fit is to send one prompt to several of them and compare. That is what Whizi is for: one subscription that includes GPT, Claude, Gemini, and image models in one workspace, instead of paying each provider separately. Start a trial, run your own four prompts, and keep whichever answer you would have been willing to send.