What is artificial intelligence, really?

A word that is everywhere

It is hard to get through a week without hearing about AI. It is in the news, in adverts, in the apps on your phone. But if someone stopped you and asked what artificial intelligence actually is, you might find it harder to answer than you expected.

Try a few quick questions. Does a rock have intelligence? Clearly not. Does a dog? In its own way, yes, since it perceives, learns and decides. Does a search engine, when it guesses what you are looking for and ranks millions of pages in an instant? That one is harder to call. The word "intelligence" turns out to be slippery, and "artificial" only makes it more so.

Rather than chase a perfect definition, it helps to look at AI from two angles that people in the field actually use. Each one tells you something true about what the technology is for.

The agent and the tool

The first way to see AI is as the art and science of building intelligent agents. An agent is something that senses its surroundings and acts within them. The inspiration here is us. As people, we perceive the world through sight and sound, we act in it with our hands, we talk to one another, we hold knowledge such as the capital of your country or how to ride a bike, and from that knowledge we reason and decide. We also learn and change over time. Each of those human abilities has a matching branch of AI: seeing becomes computer vision, moving becomes robotics, talking becomes language processing, adapting becomes machine learning. On this view, the goal is to build something that behaves, in some small corner, a little like a thinking creature.

The second way is plainer. AI is simply a set of tools. On this view you are in the business of solving real problems, sorting photos, spotting fraud, translating a menu, and AI just happens to be a useful means to that end. There is no dream of a mind, only a job that gets done better.

Neither view is the whole truth, and you do not have to choose. The tool view keeps you honest about what today's systems really do. The agent view reminds you what people have long hoped they might one day become.

Narrow today, general tomorrow?

That hope has a name. A machine that could do everything a human can do, hold a real conversation, cross from one kind of problem to a completely different one, learn any skill, would be a general artificial intelligence. We do not have one.

What we do have, and have in abundance, is narrow AI. Each system is very good at one specific thing and helpless outside it. The program that recommends your next song cannot drive a car. The one that reads your handwriting cannot play chess. Every impressive AI you meet today is narrow, however clever it seems. Whether the narrow pieces will ever add up to a general mind is one of the biggest open questions in the field.

Where this split stops being reliable

Narrow and general are a useful pair, and for most of the AI around you the split does its job. It has one soft edge worth knowing about, because you meet it the moment you use a modern chat assistant.

The examples above suggest each system is helpless outside its one task: the song recommender cannot drive, the handwriting reader cannot play chess. That is still true of most systems in use. But a newer kind of model, trained on very broad data rather than for one job, will translate a menu, summarise a document, write a short program and describe a photograph, all from the same model. People in the field call these general-purpose or foundation models, and they sit awkwardly between the two boxes: far broader than narrow AI, and nowhere near a general intelligence.

The honest word for what they can do is jagged. The same model can score well on a hard professional exam and then fail a simple question a child would answer, state something false with complete confidence, and come apart when a real situation does not match what it was trained on. So treat narrow and general as a good first map rather than the territory, and keep breadth separate from reliability in your head. A system being able to attempt many things is not evidence that it does any of them well.

Where you already meet it

For all the grand questions, most AI earns its keep quietly. It suggests the next word as you type and filters junk out of your inbox. It picks the films and songs put in front of you. It reads the numbers on a cheque and the text in a photo. Further afield, it helps read satellite images to spot signs of hardship, catches crop disease early from pictures taken by phones and drones, and watches over patients in hospital rooms.

None of these is a general mind. Each is a narrow tool, pointed at one job and doing it well. That is the honest shape of AI today: not a single thinking machine, but a growing toolbox, remarkable in places, limited in others, and already woven quietly through your ordinary day.

Prefer to see it explained? Here is the recorded Code for Albania lecture on this topic.

Check yourself

• This lesson describes two ways of looking at AI, one as an agent and one as a tool. In your own words, what is the difference, and why might the tool view keep you more honest about what today's systems actually do?
• What is the difference between narrow and general artificial intelligence, and which kind is every impressive AI you meet today?
• The lesson says most AI earns its keep quietly. Give one everyday example it mentions, and explain why it counts as a narrow tool rather than a general mind.