Artificial intelligence is software that learns from data instead of following rules a person wrote.

For most of computing history, software did exactly what it was told. A programmer wrote the rules — if this, then that — and the machine followed them. Artificial intelligence inverts that arrangement. Instead of being told the rules, the system is shown examples and works out the rules for itself. Show it enough photographs labelled cat and not a cat, and it learns to recognise a cat it has never seen. Nobody wrote down what a cat looks like; the pattern was learned.

The shift that defines it

This is the single idea underneath everything else. Traditional software is a set of instructions. AI is a set of instructions for learning the instructions. That indirection is what makes it powerful and what makes it strange: the resulting system can handle situations its creators never anticipated, but it cannot easily explain why it did what it did, because no one wrote the rule it followed.

The trade is deliberate. For narrow, well-understood problems — payroll, tax, a thermostat — written rules are cheaper, faster, and auditable, and AI adds nothing. AI earns its place precisely where the rules are too many, too subtle, or too fluid to write by hand: language, images, fraud, taste, judgement under noise.

AI, machine learning, and deep learning

These three words are often used interchangeably, and they are not the same thing. Artificial intelligence is the broad goal: software that performs tasks we associate with human intelligence. Machine learning is the method that dominates the field today — learning those tasks from data rather than explicit rules. Deep learning is one family of machine-learning techniques, built from many-layered neural networks, and it is responsible for nearly every breakthrough of the last decade. Nested, not synonymous: deep learning is a kind of machine learning, which is a way of doing AI.

What today’s AI actually does

The systems drawing attention now — the ones that write, summarise, code, and converse — are overwhelmingly large language models. Underneath the fluency, they do something narrow and mechanical: given a stretch of text, they predict what comes next, one fragment at a time. Trained on a large share of the written internet, that single skill turns out to be enough to draft an email, translate a contract, or explain a concept. Capability here is an emergent property of scale, not a set of features someone designed.

What it is not

It helps to be precise about the limits. Today’s AI does not understand in the human sense, hold beliefs, or possess goals of its own. It has no memory between conversations unless one is engineered around it, and no access to the world beyond the data it was given or the tools it is handed. Its confident tone is a property of the writing, not a signal of accuracy — which is why it can state a falsehood as smoothly as a fact. Treating fluency as truth is the most common and most expensive mistake organisations make with it.

Why it matters now

None of the core ideas are new; the mathematics is decades old. What changed is that data, computing power, and technique crossed a threshold together, and systems that were academic curiosities became useful enough to build a business on. For a decision-maker, the practical question is rarely “how does it work” and almost always “where does learning-from-data beat writing-the- rules, and where does it not.” The rest of this reference is an attempt to make that question answerable.