Artificial intelligence

Field named

Dartmouth summer research project, Hanover, New Hampshire, 1956^[3]

Share of people using generative AI

53%, within three years of its mass arrival (Stanford AI Index estimate)^[1]

Weekly users of leading AI systems

At least 700 million^[8]

Global corporate AI investment

$581.7 billion in 2025, up 130%^[1]

AI data center power capacity

29.6 gigawatts^[1]

All data center electricity use

About 415 terawatt-hours in 2024, roughly 1.5% of the world's electricity; projected near 945 by 2030^[10]

Main law in force

EU AI Act, applicable from Aug 2, 2026, with high-risk rules delayed to 2027 and 2028^[9]

What the term covers

There is no single agreed meaning of intelligence, so regulators define the technology by what it does. The definition revised by the OECD in 2023 describes a machine-based system that works from its input toward explicit or implied objectives and infers how to produce outputs: predictions, recommendations, decisions or content. It adds that such systems differ in how autonomous they are and in whether they keep adapting once deployed.^[7] That wording is broad enough to include a spam filter as well as a chatbot, and the European Union's law places both of those under the same definition while regulating them very differently.^[9]

History

In 1950 the British mathematician Alan Turing asked in a journal paper whether machines could think, and offered a practical substitute for the question: an "imitation game" now known as the Turing test.^[2] Six years later John McCarthy organized a summer meeting at Dartmouth College, built on the conjecture that learning and other features of intelligence could be described precisely enough for a machine to simulate them. Dartmouth regards the meeting as the start of AI as a research field.^[3]

Early work split between programs that manipulated symbols and rules and networks that learned from examples. Frank Rosenblatt's Perceptron of 1957 was an early learning network, limited to problems that a straight line could separate. Joseph Weizenbaum's ELIZA, from the mid-1960s, imitated conversation convincingly enough that people credited it with feelings, to his surprise. Rule-based expert systems followed in chemistry and medicine, although Stanford's MYCIN, which recommended antibiotics, never entered clinical use.^[2]

Promises outran results more than once. A critical 1973 report by James Lighthill to a research council in the United Kingdom contributed to reduced funding and a period later called the first AI winter, and a second downturn followed in the late 1980s when investors lost patience.^[2] Progress continued underneath: in 1986 David Rumelhart, Geoffrey Hinton and Ronald Williams popularized backpropagation, a way of training networks with several layers.^[2]

Public milestones then came through games. IBM's Deep Blue beat the world chess champion Garry Kasparov 3.5 to 2.5 in May 1997, mostly by brute force: 32 processors weighing about 200 million positions a second.^[4] IBM Watson won the quiz show Jeopardy! in 2011.^[2] In March 2016, AlphaGo beat Lee Sedol 4 to 1 at Go in Seoul with a different recipe, using neural networks to choose and judge moves and improving by playing against earlier versions of itself.^[5]

The current era is usually dated from 2012, when AlexNet, a network from the University of Toronto, cut the error rate in the ImageNet image recognition contest from about 25% to 16%.^[2] In June 2017, Google researchers proposed the Transformer, a design that relies on a mechanism called attention in place of older recurrent and convolutional structures, and that could be trained faster and in parallel.^[6] OpenAI's GPT-3 of 2020 applied that kind of model at the scale of 175 billion parameters, and ChatGPT, released as a free research preview in 2022, brought generative AI to a mass audience.^[2]

How it works

Most systems in wide use today are not programmed with rules for their task. Developers set up a network with many adjustable numbers, called parameters or weights, show it examples and repeatedly adjust the numbers to reduce its errors, the job backpropagation does.^[2] Reinforcement learning, by contrast, improves a system through trial and reward, as AlphaGo did against itself.^[5] Language models of the Transformer family learn from text which pieces relate to which others across a sequence, and then generate output piece by piece.^[6]

3:21 Robots that were told only to score 60 Minutes shows soccer robots that learned the game by trial and error, and explains how a network strengthens the connections that worked.

6:17 A large language model, explained An IBM Technology presenter compares large language models to autocomplete that predicts far more than the next word.

Results have improved quickly but unevenly. The 2026 International AI Safety Report, written by more than 100 experts chaired by Yoshua Bengio, says leading systems reached gold-medal results on International Mathematical Olympiad questions, and that the length of software task agents can complete has been doubling about every seven months. The same report notes that one study found 79% accuracy on questions about U.S. culture and 12% on Ethiopian culture.^[8] Stanford's AI Index reports agents succeeding on 77.3% of a benchmark of real-world tasks, up from 20% a year earlier, and robots managing only 12% of household tasks.^[1]

Adoption and economic weight

Stanford estimates that generative AI reached 53% of the population within three years, a figure that hides wide gaps: 61% in Singapore, 54% in the United Arab Emirates and 28.3% in the United States, which ranks 24th.^[1] The safety report says adoption probably remains below 10% across much of Africa, Asia and Latin America.^[8] Alphabet said in October 2026 that its Gemini app had passed 1 billion monthly active users.^[15]

Corporate AI investment reached $581.7 billion worldwide in 2025 according to the AI Index, and private investment in the United States, at $285.9 billion, was 23.1 times that in China. China leads in publications, citations, patents and industrial robots, and the lead among top models has changed hands several times since early 2025; in March 2026 a model from Anthropic was ahead by 2.7%.^[1] The number of AI scholars moving to the United States has fallen 89% since 2017.^[1]

Researchers recognized the field's reach in 2024, when the Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton for work enabling machine learning with neural networks, and half of the chemistry prize went to Demis Hassabis and John Jumper for predicting protein structures.^[14] The AI Index counts a 26% to 28% yearly rise in AI-related publications in the natural, physical and life sciences.^[1]

Jobs

Evidence on employment is still mixed. The safety report estimates that around 60% of jobs in advanced economies and 40% in emerging ones are likely to be affected, and cites studies in the United States and Denmark that found no overall effect so far, alongside declining employment for early-career workers in the most exposed occupations since late 2022.^[8] The AI Index reports that employment among software developers aged 22 to 25 has fallen nearly 20% since 2024.^[1]

10:58 Hinton on the people machines may replace Hinton names a class of people left unemployed as a risk, alongside fake news and bias in hiring and policing.

Risks and controversy

A one-sentence statement published in 2023 and signed by Hinton, Bengio and the heads of several leading laboratories said that reducing the risk of extinction from AI should rank with pandemics and nuclear war as a global priority.^[12] Researchers remain divided on how likely a loss of control is, and the safety report says models sometimes recognize when they are being tested, which makes tests before release a weaker guide to behavior afterward.^[8]

0:48 Geoffrey Hinton: does humanity know what it is doing? Asked by 60 Minutes in 2023, Hinton answers no, and says we may soon have things more intelligent than us.

Nearer harms are better documented. In studies gathered by the safety report, people mistook AI-written text for a human's 77% of the time and AI voices for real speakers 80% of the time, and one estimate put 96% of deepfake videos online as pornographic. An AI agent placed in the top 5% of more than 400 mostly human teams in a software security contest, though fully autonomous attacks had not been documented.^[8] Companies have also become less open: the average score on the Foundation Model Transparency Index fell from 58 to 40.^[1]

Physical costs draw attention too. The International Energy Agency put all data centers, not only those running AI, at about 415 terawatt-hours in 2024 and projected about 945 by 2030, slightly more than Japan uses today.^[10] Stanford reports estimated training emissions of 72,816 tons of carbon dioxide equivalent for one model, Grok 4.^[1]

Copyright is being settled in court. On July 20, 2026, a federal judge gave final approval to a $1.5 billion settlement between Anthropic and authors over pirated books used in building its models, about $3,000 per work; it covers only past acquisition and copying, and leaves claims about model outputs open.^[11] Other suits against AI developers were still pending.^[11]

Public opinion is split as well. Across surveyed countries, 59% of people feel optimistic about AI's benefits and 52% feel nervous, and only 31% of the U.S. public trusts its government to regulate it, the lowest share measured.^[1]

Regulation

The European Union's AI Act, in force since Aug 1, 2024, is the most comprehensive law so far. It sorts systems by risk: banned practices such as social scoring, high-risk uses under strict duties, transparency duties for chatbots and deepfakes, and few rules for minimal-risk uses such as spam filters. Bans began applying on Feb 2, 2025, and rules for general-purpose models on Aug 2, 2025. An amending package that took effect on July 27, 2026, delayed high-risk rules to Dec 2, 2027, and Aug 2, 2028, and added a ban on systems that generate nonconsensual sexual content.^[9] In the United States, the National Institute of Standards and Technology's AI Risk Management Framework, released on Jan 26, 2023, is voluntary.^[13]

Timeline

1. 1950 Alan Turing proposes the imitation game.^[2]
2. 1956 The Dartmouth summer research project takes place.^[3]
3. 1973 The Lighthill report criticizes AI progress and funding falls.^[2]
4. May 1997 Deep Blue beats Garry Kasparov at chess.^[4]
5. 2012 AlexNet wins attention with its ImageNet error rate.^[2]
6. March 2016 AlphaGo beats Lee Sedol at Go.^[5]
7. June 12, 2017 The Transformer paper is posted.^[6]
8. 2022 ChatGPT is released as a research preview.^[2]
9. 2024 Nobel Prizes in physics and chemistry go to AI-related work.^[14]
10. Aug 1, 2024 The EU AI Act enters into force.^[9]
11. July 20, 2026 A $1.5 billion copyright settlement between Anthropic and authors receives final approval.^[11]
12. Aug 2, 2026 The EU AI Act becomes generally applicable, with exceptions.^[9]