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. The name "artificial intelligence" comes from a 1955 proposal for a summer project at Dartmouth College in New Hampshire. An AI program does not think the way a person does. Now the question is how anyone can tell how good a given AI really is.
A neural network is a kind of AI program made of many simple parts, called nodes, joined together. It is modeled loosely on the human brain. Most of today's neural networks are organized into layers of nodes. When a neural network is trained, the numbers that set how strongly its parts connect are adjusted until it gives the right answers for the training examples. In supervised learning, a program is trained on examples that people have already labeled, while in unsupervised learning it looks for patterns in data that has no labels. In reinforcement learning, a machine learns by trial and error, guided by a reward system.
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. Machine learning works best when there is a lot of data, such as thousands or millions of examples. A spam filter can wrongly flag a good email as spam, which is called a false positive. It can also trade one kind of mistake for another: lowering its false positives can raise its false negatives.
One famous test of a machine is a game. 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.
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, 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, which is one limit of tests based on conversation. Questions about the test come in two kinds: whether any computer can really pass it, and whether passing it would show thought or intelligence. In 2014 a program was said to have passed after it fooled 33 percent of judges, but the trial was very small and the result was not reliable.
A chess computer called Deep Blue followed a 1996 match in which the world chess champion beat the computer 4 games to 2. In May 1997 Deep Blue beat the champion in a six-game match, 3½ to 2½, the first time a computer beat a reigning world chess champion in a match under tournament conditions. Some game-playing programs work by brute force, trying 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 researcher, Alexander Kronrod, called chess the Drosophila of AI, comparing it to the fruit fly that geneticists use to study inheritance.
Go is hard for brute-force methods because there are so many possibilities that looking more than a move or two ahead would need an almost infinite number of outcomes to be checked. A 2016 paper in the journal Nature described AlphaGo defeating the European Go champion 5 games to 0, a feat previously thought to be at least a decade away. Its approach used value networks to evaluate board positions and policy networks to select moves. AlphaGo fine-tuned itself by playing games against slightly changed copies of itself. In March 2016 it beat a top Go player four games to one.
Now consider language models. A large language model is trained on text from sources across the internet, such as books and web forum discussions. Large language models learn to predict the next word through pre-training: the model produces probabilities for possible next words from patterns in its training data, and one of the highest-probability words is picked to continue the text. 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. 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 often did not recognize their own errors.
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.
If the examples an AI learns from leave some people out, the AI can work worse for those 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. 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, but for darker-skinned women it was more than 20 percent in one program and more than 34 percent in the other two. NIST 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. The Act sets no rules for AI judged to be minimal or no risk, and it requires that people be told when needed to keep trust, such as when they are dealing with AI. AI used in education to decide who gets in or how a person's career goes is classed as high risk. 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.
Training and running big AI programs needs a lot of electricity. A 2024 U.S. Department of Energy report said data center load growth had tripled over the past decade, and 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. In early 2025, researchers estimated that a query to a chatbot such as ChatGPT uses about five times more electricity than a simple web search. Water is also needed to cool the hardware that trains and runs generative AI, which can strain city water supplies and local ecosystems.
People who study AI do not agree on where it is headed. A March 2025 report from the Association for the Advancement of Artificial Intelligence included a survey of 475 people in its community: 76 percent said that scaling up current AI approaches is unlikely or very unlikely to lead to general intelligence. Most also disagreed that factuality or trustworthiness would soon be solved. People should check what an AI tells them before they trust it. If a chatbot names a book or article, a person should search to confirm that it really exists, and an AI should not be asked to check its own work.
Words to know
- hallucination: confident mistake
- nodes: simple parts
- saturated: no longer challenging
- jailbreak: trick prompts
- neural: brain-like
