What AI actually is

The short version

When people say "AI" today, they almost always mean a chatbot like Claude, ChatGPT or Gemini. Under the hood, those are a kind of program called a large language model. The name sounds heavy. The idea is not.

A language model is a program that has read an enormous amount of writing and learned, from all of it, what word tends to come next. Ask it a question and it writes an answer one small piece at a time, each piece being its best guess at what should follow. It does this so well, across so many subjects, that the result reads like a thoughtful person wrote it.

That is the whole trick. Everything else follows from it.

Why it seems so smart

The model was trained on more text than any human could read in a thousand lifetimes: books, articles, websites, code, conversations. It did not memorise them. It picked up the patterns in them. How explanations are structured. How a recipe is laid out. How a polite email sounds. How a maths proof moves from one line to the next.

So when you ask it to explain a mortgage, it is not looking one up. It is producing an explanation shaped like all the good explanations it has seen. That is why it can talk about almost anything, and why it can talk about it well.

Why it gets things wrong

The same trick has a catch. The model is producing text that looks right. Most of the time, text that looks right is right. Sometimes it is not.

Ask it for a book that does not exist and it may invent a plausible title and author, because that is what an answer to that question looks like. Ask it for a phone number and it may make one up that has the right shape. People call this "hallucination". A better name is confident guessing. It is not lying. It has no idea it is wrong, because it has no separate place where the truth is kept. It only has patterns.

This is the single most important thing to understand about AI. It is brilliant at shape and unreliable at facts it cannot check. The good news is that you can give it ways to check: a web search, a document, a connection to your calendar. When it can look, it looks. When it cannot, it guesses. A well written agent is told, in plain words, to say "I could not find that" instead of guessing. Every agent built on this site carries that rule.

What it is not

It is not conscious. It does not want things. It does not remember you between conversations unless the app around it saves a note. It is not connected to the internet unless it has been given a search tool. It cannot see your files, your email or your bank unless you connect them, on purpose, and you can disconnect them in a click.

It is also not one thing. There are many models, from many companies, in many sizes, and they get better every few months. The one you used last year is not the one you would use today.

Where the word "AI" gets confusing

"AI" is also used for older, narrower things: the program that beat the world chess champion, the filter that sorts spam, the system that recommends the next video. Those are real and useful, but they each do one job. The new thing about language models is that one model can do thousands of jobs, badly to brilliantly, depending on how clearly it is asked.

That is why it is worth learning to ask well. It is the difference between a tool that wastes your afternoon and one that gives you your afternoon back.

What to remember

• A language model predicts the next word, extremely well, on almost any subject.
• It is great at shape and structure, and unreliable on facts it cannot check.
• Give it a way to check, and tell it to admit when it cannot. That is most of what makes an agent trustworthy.
• It has no memory, no internet and no access to your life unless you give it those things.