What is artificial intelligence? Machines that learn, explained

Artificial intelligence is the project of getting computers to do things that would count as intelligence if a person did them: recognise a face, translate a language, answer a question, drive a car. The phrase is deliberately broad. Most of what is called AI today is one family of techniques, machine learning, applied on a very large scale.

What AI actually means

The term covers two very different things, and keeping them apart is the fastest way to understand the field. Narrow AI is a system that does one job well: recommending a film, spotting a tumour on a scan, or predicting the next word in a sentence. General AI, usually called artificial general intelligence or AGI, would be a system that can learn and reason across any task a person can. Narrow AI exists and is in use everywhere. AGI does not exist.

Within narrow AI there is a second split that matters more than the jargon. A traditional program follows rules a person wrote. A machine learning system is given examples and works out the rules itself. Almost every recent advance is of the second kind, and that is why the field changed so quickly once there was enough data and computing power to train at scale.

How it started

The founding question was Alan Turing’s. In a 1950 paper, Computing Machinery and Intelligence, he asked whether machines can think, then replaced that unanswerable question with a practical test: if a machine’s replies cannot be told apart from a person’s, on what basis do we deny it intelligence? The test is still argued about, and it is no longer treated as a serious measure of intelligence.

The field itself began in 1956, at a summer workshop at Dartmouth College organised by John McCarthy. It was McCarthy’s proposal for that workshop that gave us the phrase artificial intelligence. The early researchers were confident to the point of recklessness; many believed a machine as capable as a person was a generation away. It was not, and the decades that followed included two long funding collapses now known as the AI winters.

What changed was not one breakthrough but a combination: far more data, far faster hardware, and a technique called deep learning that could make use of both.

How it works today

Most modern AI is machine learning. You show the system a very large number of examples, it adjusts itself to reduce its errors, and it ends up able to handle examples it has not seen before. Under that label sit a few broad approaches: supervised learning, where the examples are labelled; unsupervised learning, where the system hunts for structure on its own; and reinforcement learning, where a system learns by trial and error against a reward.

Deep learning, the approach behind the recent wave, uses artificial neural networks: layers of simple connected units that together can represent very complicated patterns. This is what powers image recognition, speech recognition, translation and the large language models that generate text and code.

The important thing to understand about all of them is that they are pattern finders. They are extremely good at finding structure in data and extremely bad at knowing when they do not know something. A language model will produce a fluent, confident answer that is simply wrong, because fluency and accuracy are not the same thing.

What it is good at, and what it is not

AI is genuinely better than people at some tasks. It can search a billion images for a match, translate between languages in real time, spot patterns in medical scans, and generate plausible text at a speed no person can match. In narrow, well-defined tasks with plenty of data, it wins.

It is unreliable in ways that are easy to miss. It has no understanding of the world in the way a person does. It cannot explain its reasoning in a way you can audit. It inherits the biases present in the data it was trained on. It is confident when it is wrong. And it is very good at tasks that resemble its training data and poor at tasks that do not, which makes it hard to know in advance where it will fail.

AGI and the singularity

AGI is the goal the field is named after and has never reached: a system that can do anything a human mind can. It is the subject of genuine disagreement. Some researchers think it is decades away, some think it is much further, and some think it is a category error. Nobody has a method.

The singularity is a related idea, framed by the mathematician Vernor Vinge in a 1993 essay and popularised by Ray Kurzweil: that a machine able to improve itself would pass human capability quickly and keep improving, after which prediction becomes pointless. It is a hypothesis about a possible future, not a description of anything that has happened, and treating it as scheduled is a mistake.

The ethical and practical problems

Bias is the most immediate. A system trained on historical data learns the patterns in that data, including the unfair ones. This has already caused real harm in hiring tools, facial recognition and credit scoring, and it is hard to fix because the bias is often in the world rather than in the code.

Then there is transparency. Many of the most capable systems are, in practice, black boxes: even their builders cannot fully explain a particular decision. That is a problem anywhere a decision affects a person’s life. Privacy is a third issue, since the data that trains these systems is often personal. And there are wider questions about work, about misinformation, and about the energy these systems consume.

Governments have started to respond. The European Union’s AI Act, agreed in 2024, sorts AI uses by risk and imposes obligations accordingly. It is the first serious attempt at a general law for the technology, and it will not be the last.

AI in science fiction

Science fiction has been writing about AI for as long as the field has existed, and it has shaped what people expect. Isaac Asimov’s Three Laws of Robotics, from the 1940s, were an attempt to imagine machines that could not harm us. HAL 9000 in 2001: A Space Odyssey imagined the opposite. Blade Runner asked what happens when the machine is more human than the people hunting it, and Her and Ex Machina asked what happens when we love one.

The most useful thing the genre offers is a warning about the frame. Almost all fictional AI is general: a mind that wants things. The AI we actually have is narrow: a tool that predicts. Confusing the two leads people to fear the wrong thing and to miss the real problems, which are duller and already here.

Common misunderstandings

AI is not a single thing. It is a label for a range of techniques, most of them statistical. It is not conscious, and there is no evidence that any current system is. It does not understand what it is doing; it predicts. And it does not improve itself. Every system in use today is built and maintained by people.

Frequently asked questions

Is AI the same as machine learning?

No. Machine learning is one approach inside AI, and currently the dominant one. Deep learning is a kind of machine learning. The words are used interchangeably in marketing far more than in engineering.

What is the Turing Test?

A test proposed by Alan Turing in 1950: a person holds a text conversation with a machine and a human and tries to tell which is which. If the machine cannot be reliably distinguished, Turing suggested we should stop asking whether it thinks. It is a thought experiment about how we would know, not a definition of intelligence.

Has any AI passed the Turing Test?

Results are claimed from time to time, and they are not very meaningful. The test measures whether a system can imitate a person in a short conversation, which says little about understanding. Most researchers treat it as a historical idea rather than a benchmark.

What is AGI?

Artificial general intelligence: a system that could learn and reason across any task a person can, rather than one narrow job. It does not exist. Whether it is achievable, and how far away it might be, is genuinely disputed.

Is AI conscious?

There is no evidence that any current system is conscious, and no agreed test for it. The systems produce text that sounds thoughtful because they are built to sound thoughtful. Treating that as a mind is the most common mistake people make about AI.

What is a large language model?

A system trained on an enormous amount of text to predict what comes next. Out of that simple objective comes the ability to write, summarise, translate and answer questions. It is not looking anything up and it does not know whether what it produces is true, which is why it can be confidently wrong.