Artificial intelligence, or AI, is machines and computer programs that can imitate smart human behavior. Many AI programs learn by looking at lots of examples, such as pictures that people have labeled as dogs. People usually give an AI program the examples that it learns from. The information it learns from is called training data. An AI program usually needs a very large amount of training data to work well. Machine learning is one way to make AI: the computer finds patterns in data.
In 1950 the mathematician Alan Turing published a paper that asked, "Can machines think?" In the game he described, a person asking questions tries to tell which of two hidden players is the machine and which is the person, and the answers are typed. John McCarthy, who wrote an early FAQ on the subject, wrote that the Turing test is one-sided. A machine that passes it should be considered intelligent, he wrote, but a machine could be intelligent without knowing enough about humans to imitate one. He also wrote that some people are easily led to believe that a rather dumb program is intelligent.
In 1996 the world chess champion beat a computer 4 games to 2. Deep Blue followed in May 1997, when it beat the champion in a six-game match, 3½ to 2½. It was the first time a computer beat a reigning world champion in a match under tournament conditions. Some game-playing programs work by brute force: they try a huge number of possible moves and replies, and Deep Blue could examine about 200 million chess moves each second. McCarthy wrote that a Russian AI researcher, Alexander Kronrod, called chess the Drosophila of AI, comparing it to the fruit fly that geneticists use to study inheritance.
In March 2016 a Go-playing program called AlphaGo beat a top Go player four games to one. Before AlphaGo, many people thought it would be about a decade before computers could challenge skilled Go players. AlphaGo fine-tuned itself by playing games against slightly changed copies of itself. Computer programs can be good at one narrow job and still be unable to do a different one: a chess program would struggle with a new variant of chess. Computers can be good at some tasks that are hard for people, such as looking through huge amounts of data very fast. Some tasks are easy for a person but hard to explain to a computer, such as telling pictures of different people apart.
The first trainable neural network, the Perceptron, was shown in 1957 by a Cornell psychologist. In July 1958 a newspaper headline said it "learns by doing," and skeptics insisted it could not reshape the relationship between humans and machines. Neural networks were first proposed in 1944 and were a major research area until 1969, when, according to computer science lore, two MIT mathematicians killed them off. By the 1980s, new algorithms for networks with more than one layer brought a revival. Even after training, it is hard to say what a neural network's settings mean.
A large language model, the kind of AI behind many chatbots, is trained on text from sources across the internet, such as books and web forum discussions. It works by predicting which word is likely to come next, a bit like autocomplete but on a much larger scale. Early next-word models, called Markov models, looked back only a word or a few words, so they were not good at writing plausible text. Because a language AI picks words by probability, it can produce a false statement that sounds right. AI mistakes are sometimes called hallucinations.
AI chatbots can make up facts and even make up the sources they name. Library guides at universities warn that AI chatbots can produce fake citations that look convincing, so each source has to be looked up. A real source named by an AI may not contain the information the AI says it does, so a person has to read the source itself. An AI tool may not know about recent events or new research if that information is not in the data it learned from.
If the examples an AI learns from leave some people out, the AI can work worse for those people. In a 2018 MIT study of three commercial face-analysis programs that guessed gender, the error rate for light-skinned men was never worse than 0.8 percent. For darker-skinned women, the error rates were more than 20 percent in one program and more than 34 percent in the other two.
Historical bias happens when training data reflects unfair conditions that existed in the world at the time. Automation bias is the human tendency to favor results from automated systems over results from non-automated ones. AI can be used to make decisions that affect people, such as whether a person is admitted to a school or approved for a loan.
As of November 2026, the European Union's AI Act, adopted in 2024, sorts AI systems into four levels of risk.
Training and running big AI programs needs a lot of electricity. A national laboratory found that U.S. data centers used about 4.4 percent of all U.S. electricity in 2023. Its June 2026 update says they could use about 11.8 percent by 2030, within a range of 9.5 to 15.3 percent.
A good way to check an AI answer is to open a new tab and look the facts up on other reliable pages. This is called lateral reading. If a chatbot names a book or article, a person should search to confirm that it really exists. An AI should not be asked to check its own work: a person should look for outside sources. People should check what an AI tells them before they trust it.
Words to know
- perceptron: early neural network
- hallucinations: confident mistakes
- bias: unfair lean
- citations: named sources
- brute: try everything
