What is artificial intelligence? AI explained in plain words

Artificial intelligence (AI) is software that does tasks we usually link with human thinking: recognizing speech, translating, spotting a tumor on a scan, writing a reply. What most AI today has in common is that nobody writes its rules by hand. It learns them, by adjusting millions or billions of numbers until its outputs match the examples it was shown.

In short

• AI is a goal, not one technology: machines doing things that seem to need intelligence.
• Almost all of today's AI is machine learning, which finds patterns in data instead of following rules a programmer wrote.
• Every AI system in use is narrow: very good at the tasks it was trained for, and unreliable outside them.

Where does the term come from?

The name was chosen on 31 August 1955, in a funding proposal for a summer workshop at Dartmouth College written by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon. Their proposal set out the field's founding bet: that "every aspect of learning or any other feature of intelligence" could in principle be described precisely enough for a machine to simulate it. The workshop ran in the summer of 1956, and the field has carried the name since. Five years earlier, Alan Turing had asked whether machines can think, and proposed a test for it.

There is still no single agreed definition, because there is no agreed definition of intelligence. The working one used by many governments comes from the OECD, which in 2023 updated its wording: an AI system is a machine-based system that, from the input it receives, infers how to generate outputs "such as predictions, content, recommendations, or decisions." That is deliberately broad. A spam filter, a route planner and a chatbot all fit.

How is AI different from ordinary software?

Ordinary software follows instructions a person wrote: if the basket total is over $50, make shipping free. Early AI worked the same way. From the 1960s to the 1980s, "expert systems" encoded the knowledge of doctors or engineers as thousands of if-then rules. They worked in narrow cases and broke on anything their authors hadn't foreseen, because the world has more cases than anyone can write down.

The approach that won is machine learning. Instead of writing rules, you collect examples of inputs with the right answers and let a program adjust itself until it gets them right. The programmer writes how the system learns, not what it knows. That is why a modern AI system can recognize a face no one described to it, and also why it can be confidently wrong in ways no one programmed.

How does a machine learn? A worked example

Here is the whole idea at toy scale. Suppose you want a program to predict taxi fares, and you have three past trips: 2 km cost $7, 5 km cost $13, and 10 km cost $23. The model is a formula with two adjustable numbers, called parameters: fare = a × distance + b.

1. Guess. Start with a = 1 and b = 0. The model predicts $2, $5 and $10.
2. Measure the error. It is short by $5, $8 and $13. Squaring and averaging those misses gives one number, the loss, here about 86.
3. Nudge. Work out which direction to move a and b to make the loss smaller, and move them a little. Because every guess was too low, both go up.
4. Repeat. After enough rounds the numbers settle at a = 2 and b = 3. Now 2 km gives $7, 5 km gives $13, and 10 km gives $23. The loss is zero.

Nobody told the program that a ride costs $3 to start and $2 a kilometer. It found those numbers in the data. A model asked about a 7 km trip it never saw would answer $17. That step, from examples to a rule that covers new cases, is what "learning" means in AI.

Real systems do exactly this with far more parameters. A neural network is a very large formula made of layers of simple units, and deep learning means using many such layers. The language models behind chatbots tune billions of parameters on trillions of words, but the loop is the same: guess, measure the error, nudge, repeat.

AI, machine learning, deep learning, generative AI: what's the difference?

The terms nest inside each other, which is why they get mixed up.

Term	What it means	Example
Artificial intelligence	Any machine doing a task that seems to need intelligence, by any method	A chess program, a hand-written rule system, a chatbot
Machine learning	AI that learns its behavior from data	A spam filter trained on labeled email
Deep learning	Machine learning with many-layered neural networks	Face recognition, speech-to-text
Generative AI	Deep learning used to produce new text, images, audio or code	A chatbot, an image generator

So generative AI is one kind of deep learning, which is one kind of machine learning, which is one way of doing AI. Most AI in daily use is not generative at all: it ranks your feed, flags a fraudulent card payment, or estimates when a delivery will arrive.

What can AI do, and where does it fail?

AI is strong where there is a lot of data and a clear measure of success. It beat the world chess champion Garry Kasparov in a match in 1997, and the Go champion Lee Sedol in 2016. It transcribes speech, translates between dozens of languages and reads medical images. Stanford's AI Index counted 223 AI-enabled medical devices approved by the US Food and Drug Administration in 2023, up from 6 in 2015, and reported that 78% of organizations said they used AI in 2024, up from 55% the year before.

The failures follow from how it learns. A model knows only the patterns in its training data, so it can be thrown by cases unlike them, and it can absorb the data's biases: a hiring model trained on a company's past hires will copy whoever that company used to favor. Chatbots predict plausible text rather than look facts up, which is why they sometimes make things up in a confident tone. And a model rarely explains its answer in a way anyone can check. The US National Institute of Standards and Technology built its AI Risk Management Framework around exactly these problems: validity, bias, transparency and accountability.

Is today's AI actually intelligent?

Every system in use is narrow AI. A model that writes fluent essays cannot drive a car, and a model that plays Go at superhuman level cannot hold a conversation. Artificial general intelligence, a system that can learn any intellectual task a person can, does not exist yet.

Still argued

Whether large language models are a path to general intelligence is one of the field's live disputes. One camp points to abilities that appeared as models grew larger, such as solving problems they were never trained on, and expects scaling to continue. Another, including researchers who built the deep-learning methods themselves, argues that predicting text is not the same as understanding the world, and that new ideas will be needed. Neither side has a test both accept, and nobody can say when, or whether, general AI will arrive.

A related argument is about risk. Some researchers think a system far more capable than people could be impossible to control; others think that worry distracts from harms already happening, such as bias and misinformation. Both views are set out in is AI an existential risk?

What should you read next?

Most readers' next question is about the chatbots. How ChatGPT works walks through how a model trained to predict the next word becomes an assistant, and what a large language model is explains the models underneath. For how the field got here, from Dartmouth through two "AI winters" of lost funding to today, read the history of artificial intelligence.

Check yourself

• Without looking back: what is the difference between a program that follows rules and one that learns them?
• Using the taxi example, explain what the loss is and what training does to it.
• Why is generative AI a kind of machine learning, but not all machine learning generative?

Sources

• A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence — John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, 1955
• Updates to the OECD's definition of an AI system explained — OECD.AI, 2023
• The 2025 AI Index Report — Stanford Institute for Human-Centered Artificial Intelligence, 2025
• AI Risk Management Framework — U.S. National Institute of Standards and Technology, 2023
• What is artificial intelligence (AI)? — IBM