What AI is (and isn't)

Four words, one set of nested circles

Ask five people what artificial intelligence means and you will get five answers, most shaped by whatever product they used last week. That is a feature of the term. Artificial intelligence has always been a goal, not a technique: the ambition to build machines that do things we would call intelligent if a person did them.

The most useful mental model is a set of nested circles. The outer circle is artificial intelligence, the whole ambition, and it includes approaches that involve no learning at all. A chess program built from hand-written rules and brute-force search is AI in the classic sense, and the machine that beat the world champion in 1997 was largely of that kind. Inside that circle sits machine learning: instead of writing the rules yourself, you give a system examples and let it work out the rules from the data. Inside machine learning sits deep learning, a family of learning methods built from many layers of simple numerical units, loosely inspired by neurons, that proved unusually good at messy inputs like images, audio and language. And inside deep learning sits generative AI, systems trained to produce new content in the same form as their training data, where earlier systems only labelled or predicted: text, images, code, sound.

Every generative system is a deep learning system, every deep learning system is a machine learning system, and every machine learning system is AI. The reverse is not true, and much confusion comes from using the outer word when the inner one is meant.

Rules you write versus rules you grow

The deepest line in the diagram is the one between the outer circle and the second. For most of the field's history, building an intelligent system meant a person writing down knowledge explicitly: interview an engineer, turn their expertise into if-then rules, encode them. These expert systems had real successes and a hard ceiling. They only knew what someone had managed to articulate, and much of human skill resists articulation. You can recognise a friend's handwriting but you cannot write the rule that does it.

Machine learning inverts the process. You collect examples of the input alongside the answer you want, and a learning algorithm adjusts a model until its outputs match. Nobody writes the rule for handwriting; the system grows one from thousands of samples. The method took over as soon as data and computing power became cheap enough, because it scales with data instead of with the patience of experts.

A great deal of valuable software sits in the outer circle only. A scheduling algorithm or a database search may look intelligent, yet neither learns. Calling them AI is not wrong so much as unhelpful, because it hides the question you care about: does this thing get better with more data, and what happens when the data changes?

Narrow, general, and the space in between

A second distinction runs at right angles to the first. A narrow system does one task, or a tight cluster of tasks, and nothing else. The 1997 chess machine could not play draughts. Almost every deployed AI system in history has been narrow, and narrowness is not a weakness; it is what makes a system testable, predictable and safe to rely on.

A general system would take on a wide range of intellectual tasks it was not specifically built for, roughly as a capable person can. Whether and when machines reach that bar is a live debate this course will not try to settle. What matters for a practitioner is the space in between. Large language models are unusual because they are trained on one task, predicting text, yet the skill that demands transfers to a surprising breadth of others: drafting, translating, classifying, coding. They are neither narrow in the old sense nor general in the strong sense. A useful phrase is broad but uneven: wide in coverage, wildly variable in reliability from one task to the next.

The AI effect: why the goalposts move

A curious pattern has repeated for decades. A capability is declared a milestone of machine intelligence. Machines achieve it. Almost immediately, people stop calling it AI. Character recognition, route planning, grandmaster-level chess, transcribing speech: each was once a headline achievement and each is now just software. This is the AI effect, and it is worth understanding instead of mocking.

It happens because our informal definition of intelligence is really a definition of whatever machines cannot yet do. Once a task is mechanised we can see the mechanism, and that makes it feel like mere computation, not thought. The goalposts move because the word was always pointing at the frontier.

Two consequences follow. First, do not waste energy arguing about whether something is really AI; ask what it does, how it was built and how it fails. Second, expect the AI features you deploy this year to feel unremarkable in a few years. That is the sign they worked and became infrastructure, not a failure.