What Is AI — and the Types of AI

Let's start with the plainest definition. Artificial intelligence is software that learns patterns from data instead of being told every rule by a programmer. That single idea — learning from examples rather than hand-coded instructions — is what separates AI from ordinary software. Everything else is a variation on it.

Here's the difference in practice. If you wanted to block spam the old way, you'd write rules: block anything containing "free money", block anything with six exclamation marks. Spammers read your rules and route around them, so you write more rules, forever. The AI way is to show a program a million emails already labelled spam or not-spam and let it work out the patterns itself — including patterns you'd never have thought to write down. Nobody programmed the final behaviour. It came out of the data.

Now the terms you'll hear constantly, from the outside in. Machine learning is the branch of AI that learns from data — the spam filter above. Deep learning is machine learning built on large neural networks, which is what made image and speech recognition suddenly work around 2012. Generative AI is the newest slice of deep learning: instead of sorting things into buckets, it produces new text, images, code and audio. And an LLM — a Large Language Model — is a generative model trained on enormous amounts of text. It's the engine behind ChatGPT, Claude and Gemini.

So they nest, each one inside the last: AI contains machine learning, which contains deep learning, which contains generative AI, which contains LLMs. When a headline says "AI", it almost always means that innermost circle. Knowing the nesting is genuinely useful — it tells you that "we use AI" on a product page could mean anything from a decades-old statistical model to a frontier LLM.

One more split worth having: narrow versus general. Every AI system in existence today is narrow. It does one class of thing — translate, rank, write, generate images — and has no understanding of anything outside that. Artificial general intelligence, the kind that matches a human across the board, does not exist, and no one can tell you honestly when or whether it will. Treat any confident claim about that timeline as opinion, including the confident ones.

Which brings us to the two things you must hold together. Today's assistants are extraordinarily capable: they will draft, summarise, translate, explain, code and brainstorm at a standard that would have looked impossible in 2020. They are also genuinely limited in specific, knowable ways. They have no beliefs. They have no memory of you between conversations unless the product deliberately adds one. And they have no mechanism for checking whether what they just told you is true — they produce what a good answer looks like, which is usually right and sometimes confidently wrong.

The most common misconception is that an assistant looks things up. It doesn't, not by default. There's no database of facts inside it being queried. It generates an answer from patterns, the way you'd finish a sentence in your own language without consulting a dictionary. Some products bolt a real search onto the model, and when they do they'll usually show you links — that's a different mechanism, and it's why cited answers tend to be more reliable than uncited ones.

Hold capability and limitation together and you have the whole skill. Everything else in this level is detail underneath it.