A working definition, the difference between narrow and general AI, and the words people use loosely.
Artificial intelligence is the business of getting computers to do things that, when a person does them, we call intelligent: recognising a face, understanding a request, planning a route, writing a summary, deciding which of two options is better.
That definition is deliberately about behaviour. It says nothing about consciousness, understanding or feeling, because nobody can measure those in a machine, and the argument about whether they matter has run for seventy years without settling. What you can measure is whether the output is any good.
Artificial intelligence is the whole field, including approaches that do not learn at all. A chess program from 1985 that follows hand-written rules is AI.
Machine learning is the part of the field where the system improves by processing examples rather than by having rules written for it. Nearly everything built since 2012 is machine learning.
Deep learning is machine learning done with neural networks of many layers. It is a subset of a subset, and it is what people usually mean when they say "AI" today.
If someone wrote down the conditions, it is a program. If the conditions were inferred from data, it is machine learning. The difference matters because you can read a program and argue with it, and you mostly cannot do that with a model.
Every system in commercial use today sits firmly on one side of this line.
Built for one task and useless outside it. A model that reads chest X-rays cannot read a contract, book a flight or play draughts. It has no idea it is looking at a lung.
Every AI product you have used is narrow, including the ones that converse well.
A system that could take on any cognitive task a person can, transfer what it learned in one domain to another, and set its own sub-goals. This does not exist.
Serious researchers put it anywhere from a decade to never. Treat confident dates in either direction as marketing.
Large language models look general because language touches everything. But breadth of topic is not breadth of ability: the same model that discusses quantum mechanics can fail at counting letters in a word.
Each of these shows up in ordinary conversation and in a surprising number of articles.
A language model predicts likely continuations of text. That produces genuinely useful output and no inner grasp of meaning. Whether that distinction will always hold is an open argument; today it explains most of the failures you will see.
Neural networks took a loose metaphor from neuroscience in the 1940s and went their own way. A biological neuron is vastly more complicated than a weighted sum, and brains do not learn by backpropagation.
More of the same kind of data stops helping quickly. What usually helps is data that covers cases the model currently gets wrong — which is harder to collect and easier to skip.
A model reproduces the patterns in its training data, including the unfair ones, and adds the authority of a number. Objectivity has to be measured and defended, not assumed.
It replaces tasks inside those jobs — boilerplate, first drafts, initial screening. The parts that involve judgement, responsibility and sitting with a person have moved slowest.
"AI-powered" is on a great many products that contain no such thing. Three questions usually settle it.