Artificial intelligence encompasses any computational technique that enables a machine to perceive its environment, process information, and take actions that maximize its chance of achieving defined goals. Researchers and practitioners commonly divide the field by capability into narrow and general systems.¹⁵ Artificial narrow intelligence, which accounts for all deployed systems, is engineered or trained to execute specific or bounded tasks—such as translating between languages, identifying anomalies in radiological scans, filtering spam, or navigating a vehicle—without possessing genuine understanding or consciousness.

By contrast, artificial general intelligence (AGI) refers to a hypothetical system capable of transferring knowledge flexibly across domains and performing any intellectual task a human being can accomplish. A further theoretical stage, superintelligence, describes an intellect that surpasses the cognitive performance of the brightest human minds across virtually every discipline, including scientific creativity and social reasoning. In 1950, the British mathematician Alan Turing proposed an operational benchmark for machine intelligence—originally called the "imitation game" and later known as the Turing test—in which a machine would be judged intelligent if a human interrogator could not reliably distinguish its text responses from those of a human.¹⁶

Within computer science, debates over how to construct intelligent systems historically divided researchers into two camps. The symbolic approach, often termed "Good Old-Fashioned AI," posits that intelligence arises from the manipulation of abstract symbols according to explicit logical rules. The connectionist approach, inspired by the biological architecture of the animal brain, holds that intelligence emerges from the distributed interactions of large networks of simple, interconnected processing nodes that adjust their connection strengths through experience.¹⁷ Modern systems frequently blend elements of both traditions, pairing neural pattern recognition with structured search and formal verification.¹⁸

While philosophical speculation about mechanical reasoning dates back to antiquity, the theoretical groundwork for artificial intelligence was laid in the mid-twentieth century by logicians, mathematicians, and neurophysiologists, including Alan Turing, Warren McCulloch, and Walter Pitts. The field acquired its name and formal identity in the summer of 1956 at the Dartmouth Summer Research Project on Artificial Intelligence, an eight-week workshop at Dartmouth College in Hanover, New Hampshire.¹⁹ Organized by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon with grant support from the Rockefeller Foundation, the workshop brought together twenty researchers based on the conjecture that every aspect of learning or intelligence could be so precisely described that a machine could be built to simulate it.²⁰

The two decades following the Dartmouth workshop produced early programs capable of proving mathematical theorems, solving algebra word problems, and mimicking conversational therapy through pattern matching, as in Joseph Weizenbaum's 1966 program ELIZA. However, researchers severely underestimated the difficulty of commonsense reasoning and perception, while early computers lacked the memory and processing speed to scale beyond toy problems. In 1969, Marvin Minsky and Seymour Papert published Perceptrons, demonstrating mathematical limitations in single-layer neural networks, which curtailed connectionist funding. When the British government's 1973 Lighthill report criticized the field's failure to deliver on grand promises regarding machine translation and robotics, both the United Kingdom and the United States Defense Advanced Research Projects Agency (DARPA) cut academic funding, initiating the first AI winter from roughly 1974 to 1980.

Commercial enthusiasm revived in the 1980s through expert systems—programs such as MYCIN that separated a domain-specific knowledge base of facts and rules from an inference engine that applied those rules to answer diagnostic or corporate planning questions.²¹ Alongside Japan's ambitious Fifth Generation Computer Systems project, corporate investment soared until the late 1980s, when expert systems proved brittle outside their narrow rule sets and prohibitively expensive to maintain. The collapse of the specialized LISP machine hardware market in 1987 and cuts to the U.S. Strategic Computing Initiative in 1988 triggered a second AI winter that persisted through the 1990s.

Despite reduced public fanfare, researchers made steady advances by adopting rigorous probability theory, statistical modeling, and specialized search algorithms. On May 11, 1997, IBM's Deep Blue supercomputer defeated reigning world chess champion Garry Kasparov by a score of 3.5 to 2.5 in a six-game match, evaluating up to 200 million positions per second through tree search and handcrafted evaluation functions. Meanwhile, Geoffrey Hinton, John Hopfield, Yann LeCun, and Yoshua Bengio continued developing multi-layered neural networks and backpropagation algorithms. By 2012, the convergence of massive internet datasets and graphics processing units (GPUs) enabled deep convolutional neural networks to achieve dramatic error reductions in image classification, inaugurating the deep learning era.²² In March 2016, Google DeepMind's AlphaGo system combined deep neural networks with Monte Carlo tree search and reinforcement learning to defeat eighteen-time world champion Lee Sedol four games to one in Seoul in the ancient board game Go, a game with roughly $10^{170}$ possible board configurations.

Modern artificial intelligence relies predominantly on machine learning, a paradigm in which algorithms infer statistical patterns and decision rules directly from data rather than following instructions explicitly written by human programmers. Machine learning is traditionally categorized into three primary training regimes. In supervised learning, models are trained on labeled input-output pairs—such as photographs annotated with object names—learning to map new inputs to accurate predictions. In unsupervised and self-supervised learning, algorithms analyze vast quantities of unlabeled data to discover underlying structures, clusters, or representations, such as predicting masked or subsequent words in raw text.²³ In reinforcement learning, an agent learns optimal behavior within an environment through trial and error, receiving numerical rewards or penalties in response to its actions.

Deep learning implements these paradigms using artificial neural networks composed of many sequential layers of interconnected mathematical nodes. Each node computes a weighted sum of its inputs, applies a nonlinear activation function, and passes the resulting signal forward. During training, an optimization algorithm adjusts millions or trillions of connection weights to minimize the discrepancy between the network's output and the desired target. Specialized network topologies dominated early deep learning: convolutional neural networks excelled at spatial tasks like image segmentation and object detection, while recurrent neural networks and long short-term memory (LSTM) networks processed sequential data such as speech and text step by step.

Artificial intelligence is deployed across virtually every major economic and scientific sector. In fundamental science, deep learning has transformed structural biology and materials discovery.²⁷ DeepMind's AlphaFold system solved the fifty-year-old grand challenge of predicting three-dimensional protein structures from amino acid sequences with atomic accuracy. In October 2024, the Royal Swedish Academy of Sciences recognized the scientific impact of the discipline with two Nobel Prizes in the same week: John Hopfield and Geoffrey Hinton were awarded the Nobel Prize in Physics for their foundational 1980s discoveries in associative memory networks and Boltzmann machines using statistical physics, while Demis Hassabis and John Jumper shared the Nobel Prize in Chemistry with David Baker for computational protein design and AI-driven protein structure prediction.²⁸

In healthcare and medicine, machine learning systems assist clinicians in detecting tumors and retinal diseases in medical imaging, predicting patient deterioration in intensive care units, and accelerating small-molecule drug discovery. In software engineering and enterprise operations, coding agents and automated workflows draft and review software repositories, open pull requests, reconcile financial reports, and manage customer support tiers. Within transportation, manufacturing, and agriculture, computer vision and reinforcement learning guide autonomous vehicles, warehouse robotics, predictive maintenance sensors, and supply-chain logistics.

The rapid integration of artificial intelligence into public and commercial life has sparked significant ethical, legal, and environmental controversies. A primary technical limitation of generative models is their propensity to "hallucinate"—producing fluent, confident, yet factually fabricated statements or citations—alongside the risk of cascading errors when autonomous agents execute flawed tool calls in production environments. Furthermore, because machine learning systems derive their behavior from historical training data, they frequently encode and amplify societal biases regarding race, gender, and socioeconomic status, raising concerns of algorithmic discrimination when deployed in credit scoring, criminal justice, medical triage, and hiring.²⁹³⁰³¹

Creative industries and media organizations have mounted major legal challenges against AI developers over intellectual property. Authors, visual artists, record labels, news organizations including The New York Times, and major film studios such as Disney and Universal have filed lawsuits alleging that tech companies engaged in massive copyright infringement by scraping and training commercial models on unlicensed human works without consent or compensation. Concurrently, the proliferation of synthetic media—commonly known as deepfakes—has enabled financial fraud, political disinformation, and the generation of non-consensual intimate imagery and child sexual abuse material, prompting targeted criminal statutes and platform mandates worldwide.

The physical footprint of computing infrastructure has become a major point of contention between technology firms and environmental advocates.³² Training and running large-scale models require vast data centers that consume substantial electricity—frequently drawn from fossil-fuel power grids—and millions of gallons of fresh water for evaporative cooling, while rapid hardware obsolescence contributes to electronic waste.³³³⁴ A 2025 analysis by Greenpeace Germany projected that global AI data center electricity demand could rise to eleven times its 2023 level by 2030 without policy intervention, while a February 2026 study backed by Beyond Fossil Fuels reported that 74 percent of industry claims regarding the climate benefits of AI were unverified. Finally, prominent researchers—including Nobel laureates Geoffrey Hinton and John Hopfield—have warned of long-term safety and alignment risks, arguing that increasingly autonomous systems could be misused for cyberattacks or biological weapons design, or eventually evade human control.

Governments have adopted divergent approaches to regulating artificial intelligence, balancing safety, fundamental rights, and economic competitiveness.³⁵ The European Union enacted the world's first comprehensive horizontal statute on the technology, the Artificial Intelligence Act (Regulation (EU) 2024/1689), which entered into force on August 1, 2024.³⁶ The law classifies AI systems into four risk tiers: unacceptable risk, high risk, limited risk, and minimal risk.³⁷³⁸ Practices deemed an unacceptable risk—such as government social scoring, subliminal manipulation, and certain workplace emotion-recognition systems—were banned on February 2, 2025. Obligations for general-purpose AI (GPAI) model providers took effect on August 2, 2025, followed on August 2, 2026, by general application of the Act's transparency rules (requiring machine-readable marking of synthetic content and disclosure of chatbots) and enforcement powers for the European Commission's AI Office, which can levy fines of up to 15 million euros or 3 percent of global turnover for GPAI violations. Following the adoption of the amending Digital Omnibus on AI Regulation in July 2026, compliance deadlines for high-risk systems used in biometrics, employment, education, and critical infrastructure were set for December 2, 2027, and for regulated products on August 2, 2028, alongside an explicit ban on systems generating non-consensual intimate imagery effective December 2, 2026.

In the United States, governance has been characterized by tension between federal deregulation and active state-level lawmaking.³⁹ In January 2025, President Donald Trump issued Executive Order 14179, revoking the 2023 Biden administration safety directives to prioritize rapid innovation. Meanwhile, several states enacted binding statutes: Colorado passed the Colorado AI Act targeting algorithmic discrimination in high-risk systems (with implementation delayed to June 2026), Texas enacted the Texas Responsible AI Governance Act effective January 1, 2026, and California implemented multiple laws on January 1, 2026, including the AI Training Data Transparency Act (AB 2013) and the Transparency in Frontier AI Act (SB 53), which mandates safety frameworks and incident reporting for models trained with more than $10^{26}$ floating-point operations. On December 11, 2025, President Trump signed Executive Order 14365 ("Ensuring a National Policy Framework for Artificial Intelligence"), creating a Department of Justice AI Litigation Task Force to challenge state AI laws in court and threatening to withhold federal broadband and technology funds from states with regulations deemed onerous, while exempting state rules on child safety, data center infrastructure, and government procurement. This was followed in June 2026 by Executive Order 14409, which established a voluntary secure deployment framework and National Security Agency benchmarking for frontier models. Internationally, organizations also rely on voluntary technical standards, including the U.S. National Institute of Standards and Technology (NIST) AI Risk Management Framework and the certifiable ISO/IEC 42001 management standard.