Artificial intelligence, or AI, is machines and computer programs that can imitate smart human behavior, such as making predictions and decisions. NIST, the U.S. National Institute of Standards and Technology, describes an AI system as an engineered or machine-based system that can, for a given set of objectives, generate outputs such as predictions, recommendations, or decisions influencing real or virtual environments. John McCarthy wrote an early FAQ on the subject. He defined AI as the science and engineering of making intelligent machines, especially intelligent computer programs. He said that AI does not have to confine itself to methods that are biologically observable. He also wrote that a solid definition of intelligence that does not depend on relating it to human intelligence did not yet exist.
McCarthy wrote that there is no single yes-or-no answer to whether a machine is intelligent, because intelligence involves many mechanisms and AI has learned to make computers carry out some but not others. He wrote that most work in AI studies the problems the world presents to intelligence rather than studying people, and that AI researchers may use methods people do not use. The name "artificial intelligence" comes from a 1955 proposal for a summer project at Dartmouth College in New Hampshire. That proposal guessed that every part of learning, or any other feature of intelligence, could be described so exactly that a machine could be made to simulate it.
The Dartmouth conference of 1956 is said to be where the field began, but historians note that AI work was already going on before 1956, so the starting date is a matter of how the field is defined. The AI that began at that conference has been called dry, logic-based AI, in contrast to approaches that directly model brain circuitry.
Researchers say neural-network ideas also help neuroscience: particular network layouts have reproduced observed features of human neuroanatomy and cognition, which suggests they capture something about how the brain processes information.
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. Turing predicted that in about fifty years computers could play the game well enough that an average questioner would have no more than a 70 percent chance of picking the right answer after five minutes. McCarthy wrote that the Turing test is one-sided: a machine that passes it should be considered intelligent, but a machine could be intelligent without knowing enough about humans to imitate one. He added that some people are easily led to believe that a rather dumb program is intelligent, which is one limit of tests based on conversation. In 2014 a program was said to have passed the Turing test after it fooled 33 percent of judges, but the trial was very small and the result was not reliable.
The first trainable neural network, the Perceptron, was shown in 1957 by a Cornell psychologist, and it had one layer of adjustable weights between its input and output. In July 1958 the U.S. Office of Naval Research unveiled the perceptron, 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.
In gradient descent, a model starts with small random settings, then repeatedly moves each setting a small amount in the direction that reduces its loss, until it cannot reduce the loss any further. Backpropagation is the most common way to train a neural network, and it makes gradient descent practical for networks with many layers. Even after training, it is hard to say what a neural network's settings mean, and looking at the weights of single connections does not show how it decides. A learning AI is tested by holding some data back from training and then checking how accurate the model is on data it has not seen.
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. In 2017 researchers published the transformer, a model design based solely on attention, with no recurrence or convolutions. A transformer uses attention: for each piece of the input, it asks how much each other piece affects its meaning, which helps the model use context. Large language models learn to predict the next word through pre-training, in which the model produces probabilities for possible next words from patterns in its training data. Pre-training is only one of several key steps in making today's large language models.
NIST defines confabulation, popularly called hallucination, as the production of confidently stated but erroneous or false content, by which users may be misled or deceived. Descriptions of hallucination say it follows from how generative models work: a language model rebuilds text from patterns in its training data rather than looking facts up in a fixed store. NIST says users may believe false AI content, often because of the confident way the response is written, and then act on it or pass it on. A 2024 Stanford study of three popular language models found hallucination rates of 69 percent to 88 percent on specific legal questions, and the models did worse on harder tasks that needed a careful reading of legal ideas or texts.
Computer programs can be good at one narrow job and still be unable to do a different one. AlphaGo, a Go-playing program, beat a top Go player four games to one in March 2016. As of 2018, a philosophy encyclopedia noted that machines could win at one game and still fail at the next: AlphaGo could not read the rules of Go in plain English and write a program that plays. AlphaGo's approach used value networks to evaluate board positions and policy networks to select moves. Some game-playing programs work by brute force: they try a huge number of possible moves and replies. Go is hard for brute-force methods because looking more than a move or two ahead would need an almost infinite number of outcomes to be checked. The same encyclopedia pointed to subjective consciousness and creativity as two great challenges that AI had not yet come to grips with.
The 2026 AI Index says AI is outpacing the tests built to measure it: tests meant to stay challenging for years were saturated in months, so they stop being useful for tracking progress. A review of widely used AI tests found invalid questions in a share of items ranging from 2 percent on one math test to 42 percent on another. An AI model can win a gold medal at the International Mathematical Olympiad and still be unable to reliably tell time: on one clock-reading test the top model was right 50.6 percent of the time, and people 90.1 percent.
In a new test of telling knowledge from belief, the 2026 AI Index says hallucination rates across 26 top models ranged from 22 to 94 percent. Almost all top AI developers report results on tests of what their models can do, but reports on tests of responsible AI, such as safety and fairness, are sparse. Several top models scored well on safety tests under normal use, but safety performance dropped for all models tested when people tried deliberate jailbreak prompts. The AI Incident Database recorded 362 AI-related incidents in 2025, up from 233 in 2024.
A language model that learned from internet text will produce text that mirrors that text, including hate speech and dangerous information. NIST says that AI bias comes not only from data and software but also from human biases and from systemic biases in institutions, and its 2022 report said that to understand AI bias, people must look beyond the computing and data to the wider society in which an AI system is built and used. Historical bias happens when training data reflects unfair conditions that existed in the world at the time; a 1960s city housing dataset, for example, reflects discriminatory lending practices. Automation bias is the human tendency to favor results from automated systems over results from non-automated ones, whatever the error rates of each.
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, while the error rates for darker-skinned women were more than 20 percent in one program and more than 34 percent in the other two. The programs' error rates were consistently higher for women than for men, and for darker-skinned people than for lighter-skinned people. NIST warns that performance gaps between groups or languages can come from training data that does not represent everyone, and that this can lead to wrong guesses about how well a system works for some people. It also says it can be hard to measure the harms of AI, because some negative impacts are difficult to assess.
As of November 2026, the EU's AI Act, adopted in 2024, sorts AI systems into four levels of risk. AI systems seen as a clear threat to the safety, livelihoods and rights of people are banned, while AI judged to be minimal or no risk has no rules under the Act. AI used in education to decide who gets in or how a person's career goes is classed as high risk, and the EU says its rules for high-risk AI in areas such as biometrics, education, employment and border control will apply from 2 December 2027. NIST's AI Risk Management Framework, by contrast, is meant to be voluntary, and the OECD AI Principles, adopted in 2019 and updated in May 2024, promote AI that is innovative and trustworthy and that respects human rights and democratic values.
A March 2025 report from the Association for the Advancement of Artificial Intelligence included a survey of 475 people in its community, and 76 percent of respondents said that scaling up current AI approaches is unlikely or very unlikely to lead to general intelligence. Most also disagreed that factuality, meaning not making up answers, or trustworthiness would soon be solved. In a FAQ, McCarthy wrote that most AI researchers believe human-level intelligence will need new fundamental ideas, so no one can predict when it will come. An AI should not be asked to check its own work, and people should check what an AI tells them before they trust it.
Words to know
- transformer: attention-based model
- confabulation: made-up content
- backpropagation: training method
- neuroscience: brain science
- saturated: no longer challenging
