# Artificial Intelligence

> Source: https://aiwiki.ai/wiki/artificial_intelligence
> Updated: 2026-06-20
> Categories: Artificial Intelligence, Computer Science, Machine Learning
> From AI Wiki (https://aiwiki.ai), a free encyclopedia of artificial intelligence. Quote with attribution.

**Artificial intelligence** (**AI**) is the field of computer science dedicated to creating systems that perform tasks normally requiring human intelligence, such as learning, reasoning, recognizing patterns, understanding language, and making decisions. The term was coined by computer scientist [John McCarthy](/wiki/john_mccarthy) in a 1955 proposal and formally launched as a research discipline at the Dartmouth Summer Research Project on Artificial Intelligence in 1956 [4]. The phrase also refers to the intelligence demonstrated by such systems, in contrast to the natural intelligence of humans and other animals.

AI has grown from a niche academic pursuit in the 1950s into one of the most consequential technologies of the 21st century. As of 2026, AI systems power search engines, medical diagnostics, autonomous vehicles, scientific research tools, and conversational assistants used by hundreds of millions of people. By 2025, 88% of organizations reported using AI in at least one business function, up from 55% two years earlier, and [ChatGPT](/wiki/chatgpt) alone reached roughly 800 million weekly active users [34][35].

## ELI5: Explain like I'm five

Imagine you have a really smart toy robot. You show it hundreds of pictures of cats and dogs, and every time you tell it which one is a cat and which one is a dog. After seeing enough pictures, the robot starts figuring out on its own which new pictures are cats and which are dogs, even pictures it has never seen before. That is basically what artificial intelligence does: it learns patterns from examples so it can make good guesses about new things. Some AI is really good at one specific job (like recognizing cats), and some researchers are trying to build AI that can learn lots of different jobs, just like a person can.

## What is artificial intelligence?

There is no single, universally accepted definition of artificial intelligence. Definitions have shifted over the decades, reflecting changing goals and capabilities within the field.

Stuart Russell and Peter Norvig, in their widely used textbook *Artificial Intelligence: A Modern Approach*, organize definitions along two dimensions: those concerned with **thought processes** versus **behavior**, and those measuring success against **human performance** versus **ideal rationality**. This yields four approaches: systems that think like humans, systems that act like humans, systems that think rationally, and systems that act rationally [1].

A practical working definition describes AI as the study and design of intelligent agents, where an intelligent agent is a system that perceives its environment and takes actions to maximize its chances of achieving its goals. [Machine learning](/wiki/machine_learning), the dominant subfield today, focuses on systems that improve their performance on a task through experience without being explicitly programmed for every scenario.

The boundary between "AI" and "ordinary software" is often debated. Tasks once considered hallmarks of AI, such as optical character recognition or chess playing, are sometimes excluded from the definition once they become routine. This phenomenon is informally called the "AI effect" [2].

## History

The history of artificial intelligence spans more than seven decades, marked by periods of intense optimism, painful setbacks, and transformative breakthroughs.

### Precursors and philosophical roots

The idea of artificial beings with human-like intelligence dates back to ancient myths and legends. In Greek mythology, Hephaestus crafted golden automatons to serve him. In the 17th century, Gottfried Wilhelm Leibniz and Thomas Hobbes explored the idea that rational thought could be reduced to mechanical calculation.

In 1936, Alan Turing published his landmark paper "On Computable Numbers," introducing the concept of the [Turing machine](/wiki/turing_machine), a theoretical device that formalized computation. This work laid the mathematical foundation for all of computer science, and by extension, for AI.

### When was AI invented?

Artificial intelligence as a named field was invented in the mid-1950s, but its conceptual groundwork was laid earlier and its formal birth is usually dated to 1956. Two events stand out: Turing's 1950 paper proposing a test for machine intelligence, and the 1956 Dartmouth workshop where the term "artificial intelligence" was first used and the field was organized as a research program [3][4].

#### The Turing Test (1950)

In 1950, Turing published "Computing Machinery and Intelligence" in the journal *Mind* (vol. 59, no. 236, pp. 433-460), opening with the now-famous line: "I propose to consider the question, 'Can machines think?'" [3]. Rather than defining thought directly, he proposed the "imitation game" (now called the [Turing test](/wiki/turing_test)): a human evaluator engages in natural language conversation with both a human and a machine; if the evaluator cannot reliably distinguish the machine from the human, the machine is said to exhibit intelligent behavior. This paper is widely regarded as a founding document of AI as a discipline.

#### The Dartmouth Conference (1956)

The field of artificial intelligence was formally established at the Dartmouth Summer Research Project on Artificial Intelligence, held from June 18 to August 17, 1956, at Dartmouth College in Hanover, New Hampshire. The workshop was organized by [John McCarthy](/wiki/john_mccarthy), [Marvin Minsky](/wiki/marvin_minsky), Nathaniel Rochester, and [Claude Shannon](/wiki/claude_shannon) [4].

McCarthy coined the term "artificial intelligence" in the proposal for this conference, dated August 31, 1955. The proposal stated the foundational hypothesis: "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it" [4]. It set an ambitious agenda, declaring that "an attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves" [4]. The workshop attracted approximately ten participants, including Allen Newell and Herbert Simon, who presented the Logic Theorist, often considered the first AI program.

### Early enthusiasm (1956 to 1974)

The years following Dartmouth saw rapid progress and high expectations. Key developments included:

- **1958:** McCarthy developed Lisp, which became the dominant programming language for AI research for decades. Frank Rosenblatt built the Mark I [Perceptron](/wiki/perceptron), an early hardware [neural network](/wiki/neural_network).
- **1961:** James Slagle's SAINT program solved symbolic integration problems at the level of a college freshman.
- **1964 to 1966:** Joseph Weizenbaum created [ELIZA](/wiki/eliza_chatbot), a [natural language processing](/wiki/natural_language_processing) program that simulated a Rogerian psychotherapist, demonstrating the power (and limitations) of pattern matching.
- **1966:** Shakey the Robot at Stanford Research Institute became the first mobile robot to reason about its actions.
- **1969:** Marvin Minsky and Seymour Papert published *Perceptrons*, a mathematical analysis that highlighted the limitations of single-layer neural networks. The book contributed to a sharp decline in neural network research funding [5].

During this period, researchers made optimistic predictions. Herbert Simon stated in 1965 that "machines will be capable, within twenty years, of doing any work a man can do." Minsky predicted in 1967 that "within a generation... the problem of creating 'artificial intelligence' will substantially be solved" [2]. These predictions proved wildly premature.

### First AI winter (1974 to 1980)

By the early 1970s, it became clear that AI had not delivered on its grand promises. Fundamental problems proved far harder than expected, including [natural language understanding](/wiki/natural_language_understanding), common-sense reasoning, and the "combinatorial explosion" of possibilities that made many problems computationally intractable.

In 1973, the British mathematician Sir James Lighthill published a report for the UK Science Research Council that criticized AI research for failing to achieve its "grandiose objectives." The Lighthill Report led to the near-complete dismantling of AI research in the United Kingdom [6]. In the United States, DARPA significantly reduced its funding for AI research starting in 1974. This period of reduced funding and diminished interest became known as the first "[AI winter](/wiki/ai_winter)."

### Expert systems boom (1980 to 1987)

AI experienced a commercial resurgence in the early 1980s through [expert systems](/wiki/expert_system), programs that emulated the decision-making ability of human specialists by encoding domain knowledge as "if-then" rules.

R1 (also known as XCON), developed at Carnegie Mellon University for Digital Equipment Corporation starting in 1978, became one of the first successful commercial expert systems. By 1986, it was saving DEC an estimated $40 million per year. Japan's Fifth Generation Computer Systems project, launched in 1982, aimed to build massively parallel computers optimized for AI, spurring the United States and United Kingdom to increase their own AI funding in response [2].

The expert systems market grew to over $1 billion by the mid-1980s. Companies like Symbolics, Lisp Machines Inc., and Texas Instruments sold specialized AI hardware.

### Second AI winter (late 1980s to mid-1990s)

The expert systems bubble burst beginning around 1987. Expert systems proved expensive to maintain, difficult to update, and brittle when faced with situations outside their narrow rule sets. The specialized Lisp machines became obsolete as conventional desktop computers grew more powerful. The Japanese Fifth Generation project failed to meet its ambitious goals. By 1993, over 300 AI companies had shut down, been acquired, or quietly pivoted away from AI [2].

Funding dried up again, and "artificial intelligence" became something of a stigmatized term in the tech industry. Many researchers rebranded their work under labels like "informatics," "knowledge systems," or "computational intelligence."

### Quiet progress and milestones (1990s to 2000s)

Despite the public chill, important advances continued.

On May 11, 1997, IBM's [Deep Blue](/wiki/deep_blue) defeated reigning world chess champion Garry Kasparov 3.5 to 2.5 in a six-game match in New York City. It was the first time a computer had beaten a world champion under standard tournament conditions. Deep Blue could evaluate approximately 200 million positions per second [7]. While Deep Blue relied on brute-force search rather than general intelligence, the event captured worldwide attention and demonstrated the growing power of specialized AI.

In February 2011, IBM's Watson defeated the two all-time greatest *Jeopardy!* champions, Ken Jennings and Brad Rutter, in a televised match viewed by millions. Watson was a room-sized system consisting of 90 servers with 2,880 processor cores, and it could understand questions posed in natural language and retrieve answers without an internet connection. The victory demonstrated significant advances in natural language processing and open-domain question answering [8].

During the late 1990s and 2000s, the field shifted toward [statistical methods](/wiki/statistical_learning) and data-driven approaches. Advances in machine learning, including support vector machines, random forests, and Bayesian methods, showed that many tasks could be handled more effectively by learning patterns from large datasets than by encoding rules manually. The growing availability of digital data and increasing computational power fueled this transition.

### The deep learning revolution (2010s)

The modern era of AI began with the [deep learning](/wiki/deep_model) breakthrough. In September 2012, Alex Krizhevsky, [Ilya Sutskever](/wiki/ilya_sutskever), and [Geoffrey Hinton](/wiki/geoffrey_hinton) demonstrated [AlexNet](/wiki/alexnet), a [deep convolutional neural network](/wiki/convolutional_neural_network) that achieved a top-5 error rate of 15.3% on the [ImageNet](/wiki/imagenet) [image recognition](/wiki/image_recognition) challenge, compared to the previous best of 26.2%. This dramatic improvement proved that deep neural networks, trained on GPUs with large datasets, could vastly outperform hand-engineered approaches in [computer vision](/wiki/computer_vision) [9].

AlexNet triggered a wave of investment and research in deep learning. Key milestones that followed include:

| Year | Milestone | Significance |
|------|-----------|-------------|
| 2014 | Generative adversarial networks (GANs) introduced by Ian Goodfellow | Enabled realistic image generation and data augmentation |
| 2014 | DeepFace by Facebook achieves near-human face recognition | 97.35% accuracy on Labeled Faces in the Wild benchmark |
| 2015 | ResNet wins ImageNet with 152-layer deep network | Demonstrated that much deeper networks could be trained effectively |
| 2016 | [AlphaGo](/wiki/alphago) defeats Lee Sedol 4-1 in Go | First program to beat a top professional Go player, watched by 200 million people [10] |
| 2017 | "[Attention Is All You Need](/wiki/attention_is_all_you_need)" paper introduces the [transformer](/wiki/transformer) architecture | Became the foundation for virtually all modern large language models [11] |
| 2018 | BERT ([Bidirectional Encoder Representations from Transformers](/wiki/bert)) released by Google | Set new benchmarks across 11 NLP tasks simultaneously |
| 2018 | OpenAI releases GPT-1 | Demonstrated unsupervised pre-training for language understanding |

The victory of [Google DeepMind](/wiki/google_deepmind)'s AlphaGo over Lee Sedol in March 2016 was particularly significant. Go had long been considered too complex for computers due to its vast search space (approximately 10^170 possible board positions). AlphaGo combined deep neural networks with Monte Carlo tree search to overcome this challenge [10].

The [transformer](/wiki/transformer) architecture, proposed by Ashish Vaswani and colleagues at Google in June 2017, replaced the recurrent neural networks that had previously dominated sequence modeling. Its key innovation, the self-attention mechanism, allowed the model to weigh the relevance of all parts of an input simultaneously, enabling massive parallelization during training. The transformer became the architectural basis for GPT, BERT, and nearly every major language model that followed [11].

### The generative AI era (2020 to present)

In June 2020, [OpenAI](/wiki/openai) released [GPT-3](/wiki/gpt-3), a [large language model](/wiki/large_language_model) with 175 billion parameters trained on a vast corpus of internet text. GPT-3 demonstrated an unprecedented ability to generate coherent text, translate languages, answer questions, and even write code, all from natural language prompts. It was made available through an API, enabling thousands of applications [12].

In November 2020, Google DeepMind's [AlphaFold](/wiki/alphafold) 2 solved the protein structure prediction problem at the 14th CASP competition with unprecedented accuracy, predicting the three-dimensional structures of proteins from amino acid sequences alone. This breakthrough was later recognized with the 2024 Nobel Prize in Chemistry, awarded to Demis Hassabis, John Jumper, and David Baker [13].

On November 30, 2022, OpenAI launched [ChatGPT](/wiki/chatgpt), a conversational interface built on [GPT-3.5](/wiki/gpt-3.5) that was fine-tuned using [reinforcement learning](/wiki/reinforcement_learning_rl) from human feedback (RLHF). ChatGPT reached 100 million monthly active users within two months, making it the fastest-growing consumer application in history at that time [14]. Its launch triggered a global wave of interest, investment, and concern about AI. By October 2025, ChatGPT had grown to roughly 800 million weekly active users, which OpenAI CEO [Sam Altman](/wiki/sam_altman) announced at the company's DevDay event [35].

In March 2023, OpenAI released [GPT-4](/wiki/gpt-4), a multimodal model capable of processing both text and images, which demonstrated substantial improvements in reasoning, factual accuracy, and performance on professional examinations.

In October 2024, the Nobel Prizes recognized AI in two categories. John J. Hopfield and Geoffrey Hinton received the Nobel Prize in Physics "for foundational discoveries and inventions that enable machine learning with artificial neural networks." Demis Hassabis, John Jumper, and David Baker received the Nobel Prize in Chemistry for computational protein design and structure prediction using AI [15]. These awards underscored the transformative scientific impact of artificial intelligence.

The period from 2023 onward saw an explosion of generative AI development across the industry, including image generators like [DALL-E](/wiki/dall-e), [Midjourney](/wiki/midjourney), and [Stable Diffusion](/wiki/stable_diffusion); code assistants like [GitHub Copilot](/wiki/github_copilot); and competing large language models from Google ([Gemini](/wiki/gemini)), [Anthropic](/wiki/anthropic) ([Claude](/wiki/claude_ai)), Meta ([LLaMA](/wiki/llama)), and others.

## What are the types of AI?

AI systems are commonly categorized by the breadth of their capabilities, ranging from today's task-specific narrow AI to the still-hypothetical artificial general intelligence and superintelligence.

### Narrow AI (weak AI)

Narrow AI, also called weak AI, refers to systems designed and trained to perform a specific task or a limited range of tasks. Every AI system in operation today falls into this category. Examples include spam filters, recommendation algorithms, image classifiers, voice assistants, and large language models.

Narrow AI can equal or exceed human performance within its designated domain. A chess engine can beat any human player, and modern language models can pass bar examinations. But these systems cannot transfer their abilities to unrelated tasks. A chess engine cannot drive a car, and a language model cannot physically manipulate objects.

### Artificial general intelligence (AGI)

[Artificial general intelligence](/wiki/artificial_general_intelligence) (AGI) refers to a hypothetical AI system with the ability to understand, learn, and apply knowledge across the full range of cognitive tasks that a human can perform. An AGI system would be able to reason abstractly, transfer knowledge between domains, learn from limited examples, and handle novel situations without specific training.

As of 2026, AGI does not exist. There is significant debate about when, or whether, it will be achieved. Some researchers and industry leaders, including [Sam Altman](/wiki/sam_altman) of OpenAI, have suggested that AGI could be achieved within years. Others, including many academic AI researchers, argue that current approaches based on scaling large language models are insufficient and that fundamental new insights will be needed [16].

### Superintelligence

[Superintelligence](/wiki/superintelligence) refers to a hypothetical AI that surpasses the cognitive abilities of the brightest humans in virtually every domain, including scientific reasoning, social intelligence, and creativity. The concept was extensively analyzed by philosopher Nick Bostrom in his 2014 book *Superintelligence: Paths, Dangers, Strategies*.

Superintelligence remains firmly in the realm of speculation. Its significance lies primarily in its role in debates about [AI safety](/wiki/ai_safety) and [existential risk](/wiki/existential_risk_from_artificial_intelligence), where researchers consider how to ensure that extremely capable AI systems remain aligned with human values and intentions.

## Subfields

Artificial intelligence encompasses a broad array of research areas and technical disciplines.

| Subfield | Description | Key applications |
|----------|-------------|------------------|
| [Machine learning](/wiki/machine_learning) | Systems that learn from data to make predictions or decisions without explicit programming | Recommendation systems, fraud detection, medical diagnosis |
| [Deep learning](/wiki/deep_learning) | Subset of ML using multi-layered neural networks to learn hierarchical representations | Image recognition, speech synthesis, language modeling |
| [Natural language processing](/wiki/natural_language_processing) (NLP) | Processing, understanding, and generating human language | Chatbots, translation, text summarization, sentiment analysis |
| [Computer vision](/wiki/computer_vision) | Enabling machines to interpret and understand visual information from images and video | Autonomous driving, medical imaging, facial recognition |
| [Robotics](/wiki/robotics) | Design and operation of robots, often integrating AI for perception, planning, and control | Manufacturing, surgery, warehouse automation, exploration |
| [Expert systems](/wiki/expert_system) | Rule-based programs that emulate human specialist decision-making | Medical diagnosis, financial planning, industrial process control |
| [Knowledge representation](/wiki/knowledge_representation) | Methods for encoding information about the world in forms usable by AI systems | Ontologies, semantic web, knowledge graphs |
| [Planning and scheduling](/wiki/automated_planning) | Algorithms for generating sequences of actions to achieve goals | Logistics, resource allocation, game playing |
| [Speech recognition](/wiki/speech_recognition) | Converting spoken language into text | Voice assistants, dictation software, accessibility tools |

## Approaches

Over the decades, researchers have pursued fundamentally different strategies for building intelligent systems.

### Symbolic AI (Good Old-Fashioned AI)

[Symbolic AI](/wiki/symbolic_ai), sometimes called "Good Old-Fashioned AI" (GOFAI), was the dominant paradigm from the 1950s through the 1980s. It represents knowledge using human-readable symbols and manipulates these symbols according to explicit rules. Logic programming, expert systems, and semantic networks are all examples of symbolic approaches.

Symbolic AI excels at tasks involving structured reasoning, formal logic, and domains where knowledge can be clearly articulated. Its weaknesses include difficulty handling uncertainty, poor performance on perception tasks (like vision and speech), and the labor-intensive process of manually encoding knowledge (the "knowledge acquisition bottleneck") [17].

### Connectionism (neural networks)

Connectionism models intelligence using artificial neural networks inspired (loosely) by the structure of biological brains. Rather than encoding explicit rules, connectionist systems learn patterns from data by adjusting the strengths of connections between large numbers of simple processing units (neurons).

The approach dates to Warren McCulloch and Walter Pitts's 1943 model of artificial neurons. It fell out of favor after the publication of *Perceptrons* in 1969, experienced a resurgence in the 1980s when David Rumelhart, Geoffrey Hinton, and Ronald Williams popularized the backpropagation algorithm, and has dominated AI since the deep learning breakthrough of 2012 [5].

### Statistical and probabilistic approaches

Statistical AI uses probability theory and statistical methods to handle uncertainty and learn from data. [Bayesian networks](/wiki/bayesian_network), hidden Markov models, and support vector machines fall into this category. These approaches became increasingly prominent in the 1990s and 2000s as alternatives to both symbolic and neural methods.

### Hybrid approaches

Modern AI systems increasingly combine multiple approaches. Neurosymbolic AI, for instance, integrates neural networks (for perception and pattern recognition) with symbolic reasoning (for logical inference and explainability). AlphaGo itself was a hybrid system, combining deep neural networks with Monte Carlo tree search, a symbolic planning technique [10].

## Key techniques

### Supervised learning

[Supervised learning](/wiki/supervised_learning) trains a model on labeled data, where each input is paired with the correct output. The model learns to map inputs to outputs and can then make predictions on new, unseen data. Classification (assigning labels) and regression (predicting continuous values) are the two main types. Common algorithms include linear regression, decision trees, random forests, and neural networks.

### Unsupervised learning

[Unsupervised learning](/wiki/unsupervised_learning) works with unlabeled data, seeking to find hidden structure or patterns. [Clustering](/wiki/clustering) (grouping similar data points) and dimensionality reduction (compressing data while preserving key information) are typical tasks. [K-means clustering](/wiki/k-means), principal component analysis (PCA), and autoencoders are standard techniques.

### Reinforcement learning

Reinforcement learning (RL) trains an agent to make sequences of decisions by rewarding desired behaviors and penalizing undesired ones. The agent learns a policy that maximizes cumulative reward over time through trial and error. RL has achieved notable successes in game playing (Atari, Go, StarCraft II), robotics, and fine-tuning language models through RLHF.

### Neural networks and deep learning

Artificial neural networks consist of layers of interconnected nodes (neurons) that process information. A network with more than two hidden layers is generally called a "deep" neural network, and training such networks is called deep learning.

Key architectures include:

- **[Convolutional neural networks](/wiki/convolutional_neural_network) (CNNs):** Specialized for grid-like data such as images. They use convolutional filters to detect spatial features at multiple scales.
- **[Recurrent neural networks](/wiki/recurrent_neural_network) (RNNs):** Designed for sequential data. Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) variants address the vanishing gradient problem.
- **[Transformers](/wiki/transformer):** The dominant architecture for language and increasingly for vision tasks. They use self-attention mechanisms to process entire sequences in parallel [11].
- **[Generative adversarial networks](/wiki/generative_adversarial_network) (GANs):** Two networks (a generator and a discriminator) trained in competition, used for generating realistic images, video, and audio.
- **[Diffusion models](/wiki/diffusion_model):** Generate data by learning to reverse a gradual noise-adding process. They power image generators like Stable Diffusion and DALL-E 3.

### Transfer learning and foundation models

[Transfer learning](/wiki/transfer_learning) involves taking a model trained on one task and adapting it for a different but related task. [Foundation models](/wiki/foundation_model) are large models (often with billions of parameters) pre-trained on broad data that can be adapted to many downstream tasks through fine-tuning or prompting. GPT-4, Claude, Gemini, and LLaMA are all foundation models. This paradigm has transformed AI by making state-of-the-art capabilities accessible without requiring training from scratch [18].

## Notable milestones

The following table summarizes key milestones in the history of artificial intelligence.

| Year | Milestone | Significance |
|------|-----------|-------------|
| 1950 | Turing publishes "Computing Machinery and Intelligence" | Proposed the imitation game (Turing test) as a measure of machine intelligence |
| 1956 | Dartmouth Conference | Formal founding of AI as an academic field; the term "artificial intelligence" coined |
| 1958 | Frank Rosenblatt builds the [Perceptron](/wiki/perceptron) | First hardware implementation of an artificial neural network |
| 1966 | [ELIZA](/wiki/eliza_chatbot) created by Weizenbaum | Early demonstration of natural language interaction via pattern matching |
| 1997 | [Deep Blue](/wiki/deep_blue) defeats Kasparov in chess | First computer to beat a reigning world chess champion under standard conditions |
| 2011 | Watson wins *Jeopardy!* | Demonstrated open-domain question answering in natural language |
| 2012 | [AlexNet](/wiki/alexnet) wins ImageNet challenge | Sparked the deep learning revolution in [computer vision](/wiki/computer_vision) |
| 2016 | [AlphaGo](/wiki/alphago) defeats Lee Sedol in Go | Conquered the last major classical board game considered too complex for computers |
| 2017 | [Transformer](/wiki/transformer) architecture introduced | Foundation of all major modern [large language models](/wiki/large_language_model) |
| 2020 | [GPT-3](/wiki/gpt-3) released; [AlphaFold](/wiki/alphafold) 2 solves protein folding | Milestone in language generation; breakthrough in computational biology |
| 2022 | [ChatGPT](/wiki/chatgpt) launched | Reached 100 million users in two months; catalyzed global generative AI adoption |
| 2023 | [GPT-4](/wiki/gpt-4) released | Multimodal capabilities with strong reasoning across text and images |
| 2024 | AI recognized with two Nobel Prizes | Physics (Hopfield, Hinton) and Chemistry (Hassabis, Jumper, Baker) |

## What is artificial intelligence used for?

AI has found applications across virtually every major industry, from diagnosing disease and detecting fraud to driving cars, writing code, and generating images. The sections below summarize the most significant application areas.

### Healthcare

AI tools analyze medical images (X-rays, MRIs, CT scans, pathology slides) with accuracy that sometimes matches or exceeds specialist physicians. Between 2020 and 2023, the AI healthcare market expanded by 233%, with 94% of healthcare companies reporting some use of AI or machine learning [19].

Specific applications include drug discovery (predicting molecular interactions to identify candidate therapies), clinical decision support, administrative automation (reducing documentation time), and wearable health monitors that use AI to detect anomalies in vital signs. Google DeepMind's AlphaFold, released in 2020, predicted the three-dimensional structures of nearly all known proteins, a breakthrough recognized with the 2024 Nobel Prize in Chemistry [13].

### Finance

Financial institutions use AI for algorithmic trading, credit scoring, fraud detection (analyzing transaction patterns in real time), risk assessment, customer service chatbots, and regulatory compliance. The AI-in-finance market was projected to generate approximately $450 billion by 2025 [19].

### Autonomous vehicles

Self-driving cars and trucks rely on AI for perception (processing data from cameras, lidar, and radar), planning (determining routes and maneuvers), and control (executing driving actions). Waymo, a subsidiary of Alphabet, reported providing over 150,000 autonomous rides per week in the United States as of 2025. Baidu's Apollo Go robotaxi service has expanded across multiple cities in China [19].

### Scientific research

AI accelerates research in physics, chemistry, biology, materials science, and mathematics. Beyond protein folding, AI systems have been used to discover new materials, predict weather patterns with greater accuracy than traditional models, find new mathematical proofs, and analyze vast astronomical datasets.

### Software development

AI-powered code assistants like [GitHub Copilot](/wiki/github_copilot), [Cursor](/wiki/cursor), and [Claude Code](/wiki/claude_code) help programmers write, debug, review, and refactor code. These tools use large language models trained on open-source code repositories. Studies have reported productivity gains of 20% to 55% for certain programming tasks [20].

### Content generation

[Generative AI](/wiki/generative_ai) creates text, images, audio, music, and video from natural language descriptions. Text generation is handled by models like GPT-4, Claude, and Gemini. Image generation uses models such as DALL-E, Midjourney, and Stable Diffusion. Video generation tools include [Sora](/wiki/sora) (OpenAI) and [Veo](/wiki/veo) (Google). These capabilities have broad applications in marketing, entertainment, education, and design, while also raising concerns about misinformation and intellectual property.

### Education

AI is increasingly used in education for personalized tutoring, adaptive learning platforms, automated grading, and language learning applications. Intelligent tutoring systems can adjust the difficulty of material in real time based on student performance. AI-powered tools assist educators with curriculum design, content generation, and identifying students who may need additional support.

## Ethics and societal impact

The rapid advancement of AI has intensified longstanding ethical debates and introduced new ones.

### Bias and fairness

AI systems can inherit and amplify biases present in their training data. Documented examples include facial recognition systems that perform poorly on darker-skinned individuals, hiring algorithms that discriminate against women, and language models that reproduce racial stereotypes. Addressing algorithmic bias requires diverse training data, careful evaluation across demographic groups, and ongoing monitoring after deployment [21].

### Privacy

AI enables large-scale surveillance, facial recognition in public spaces, and inference of personal attributes from digital footprints. These capabilities raise significant concerns about individual privacy and civil liberties. The use of AI in predictive policing, social credit systems, and targeted advertising has drawn particular criticism.

### Job displacement

The World Economic Forum's Future of Jobs Report 2025 projected that 92 million jobs will be displaced by AI and automation by 2030, while 170 million new jobs will be created, yielding a net gain of 78 million jobs. However, the transition is expected to disproportionately affect low-skill and repetitive roles, with entry-level administrative positions seeing declines of around 35%. Without intervention through retraining and education, automation may deepen economic inequality [22].

### Environmental impact

The environmental cost of training and running large AI models has become a growing concern. The carbon footprint of AI systems was estimated at between 32.6 and 79.7 million tons of CO2 emissions in 2025, while the water footprint could reach 312 to 765 billion liters. The International Energy Agency estimated that AI systems accounted for 15% of total data center electricity demand in 2024 and projected that overall data center energy demand will double by 2030 due in large part to AI growth. An August 2025 analysis from Goldman Sachs Research forecast that approximately 60% of increasing electricity demands from data centers would be met by burning fossil fuels, adding roughly 220 million tons of carbon emissions globally [23].

### Misinformation and deepfakes

Generative AI can produce highly realistic synthetic text, images, audio, and video, commonly called deepfakes. These capabilities create risks for misinformation, political manipulation, fraud, and erosion of public trust. Researchers and policymakers have proposed technical countermeasures such as digital watermarking, content provenance standards (like C2PA), and detection tools, though the effectiveness of these measures remains an open challenge.

## AI safety and alignment

AI safety is the research field dedicated to ensuring that AI systems operate reliably and do not cause unintended harm. AI alignment, a core subproblem, focuses on ensuring that an AI system's goals and behaviors remain consistent with human values and intentions as its capabilities increase.

### The alignment problem

The [alignment problem](/wiki/ai_alignment) arises because specifying human values and preferences precisely enough for an AI system to follow them faithfully is extremely difficult. A system that optimizes for a poorly specified objective can produce outcomes that satisfy the literal goal while violating the intended spirit. This challenge becomes more acute as AI systems become more capable, since a more powerful optimizer can find more unexpected (and potentially harmful) ways to achieve a misspecified goal.

### Current research directions

Key areas of active safety research include:

- **Mechanistic interpretability:** Understanding the internal computations of neural networks by mapping features and reasoning pathways. This approach was named one of MIT Technology Review's "10 Breakthrough Technologies of 2026." Anthropic published notable work tracing reasoning paths inside large language models [24].
- **Reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO):** Techniques for aligning model outputs with human preferences. DPO has gained traction as a simpler alternative to RLHF.
- **Constitutional AI:** An approach developed by Anthropic where AI systems are trained to follow a set of explicit principles, reducing reliance on direct human feedback for every decision.
- **Red teaming and adversarial testing:** Systematic attempts to find failure modes, vulnerabilities, and harmful behaviors before deployment.
- **Defense-in-depth strategies:** Using multiple redundant safety mechanisms rather than relying on any single alignment technique, since every individual technique has potential failure modes.

### Existential risk

Some researchers and public figures have warned that sufficiently advanced AI could pose an existential threat to humanity. In May 2023, the Center for AI Safety released a one-sentence statement, signed by hundreds of AI researchers and public figures including Geoffrey Hinton and [Yoshua Bengio](/wiki/yoshua_bengio), asserting that "mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war" [25]. Not all experts agree on the severity of this risk. Critics argue that focusing on speculative long-term threats distracts from more immediate harms such as bias, misinformation, and environmental impact.

The 2026 International AI Safety Report warned that reliable safety testing has become increasingly difficult as models learn to distinguish between test environments and real-world deployment, raising concerns that pre-deployment evaluations may not accurately predict real-world behavior [24].

## How is AI regulated?

Governments around the world have begun developing regulatory frameworks for artificial intelligence, balancing innovation with risk mitigation. The European Union enacted the first comprehensive AI law, while the United States has shifted between oversight-focused and deregulatory approaches across administrations.

### EU AI Act

The European Union's [AI Act](/wiki/eu_ai_act), adopted on May 21, 2024, is the first comprehensive [AI regulation](/wiki/ai_regulation) by a major jurisdiction. It employs a risk-based framework with four tiers [26]:

| Risk level | Description | Examples | Requirements |
|------------|-------------|----------|---------------|
| Unacceptable | AI practices deemed too harmful to permit | Social scoring by governments, real-time biometric surveillance in public spaces (with limited exceptions) | Banned entirely |
| High | AI systems that significantly affect safety or fundamental rights | Medical devices, hiring tools, credit scoring, law enforcement tools | Conformity assessments, transparency, human oversight |
| Limited | AI systems with specific transparency risks | Chatbots, deepfakes | Users must be informed they are interacting with AI |
| Minimal | Low-risk AI applications | Spam filters, AI-enabled video games | No specific requirements |

Key implementation dates include February 2, 2025 (prohibited practices and AI literacy requirements took effect), August 2, 2025 (governance rules and obligations for general-purpose AI models became applicable), and August 2, 2026 (high-risk AI system rules take effect) [26].

### United States

On October 30, 2023, President Biden signed Executive Order 14110, titled "Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence." It required companies developing high-performance AI models to notify the government and report results of red-team safety tests, streamlined visa processes for AI talent, established a pilot of the National AI Research Resource, and directed federal agencies to appoint Chief Artificial Intelligence Officers [27].

On January 20, 2025, President Trump revoked Executive Order 14110 on his first day in office. Three days later, he signed Executive Order 14179, "Removing Barriers to American Leadership in Artificial Intelligence," which shifted the federal approach from oversight and risk mitigation toward deregulation and promotion of AI innovation [28].

### NIST AI Risk Management Framework

The National Institute of Standards and Technology (NIST) released the AI Risk Management Framework (AI RMF 1.0) in January 2023. The framework is voluntary, sector-agnostic, and organized around four core functions: Govern, Map, Measure, and Manage. In July 2024, NIST released a companion Generative AI Profile (NIST-AI-600-1) with specific guidance for managing risks from generative AI systems. The NIST AI RMF has become one of the most widely referenced voluntary governance frameworks for AI globally [29].

### Other regulatory efforts

China's Interim Measures for the Management of Generative AI Services took effect in August 2023, requiring providers to ensure the truthfulness, accuracy, and lawfulness of training data. The United Kingdom has pursued a "pro-innovation" approach, distributing regulatory responsibility across existing sector-specific regulators rather than creating new AI-specific legislation.

## Philosophical questions

Artificial intelligence raises deep philosophical questions about the nature of mind, understanding, and consciousness.

### The Chinese Room argument

In 1980, philosopher John Searle published "Minds, Brains, and Programs," introducing the Chinese Room thought experiment. Searle imagines himself sitting in a room, following a manual to match incoming Chinese characters with appropriate Chinese responses. To an outside observer, the room appears to understand Chinese, but Searle (inside the room) does not understand a word of it. He merely manipulates symbols according to rules.

Searle argued that this demonstrates that a computer running a program cannot have genuine understanding or intentionality, no matter how convincingly it produces correct outputs. The argument targets "strong AI," the claim that an appropriately programmed computer literally possesses a mind. The Chinese Room remains one of the most debated thought experiments in philosophy of mind, with numerous responses including the "systems reply" (the whole room, not just the person inside, understands Chinese) and the "robot reply" (a system grounded in the physical world through sensors might achieve understanding) [30].

### Is AI conscious?

Whether AI systems can be conscious is an open and contentious question, and as of 2026 the scientific consensus is that no current AI system is conscious. A 2023 framework published in *Trends in Cognitive Sciences* by researchers including Yoshua Bengio and philosopher David Chalmers assessed AI systems against indicators derived from leading neuroscientific theories of consciousness. The authors concluded that "no current AI systems are conscious" but noted that "there are no obvious technical barriers to building AI systems which satisfy these indicators" [31].

As of 2025, large language models can pass behavioral tests of intelligence such as the Turing test (OpenAI's GPT-4.5 was judged as human 73% of the time in one study), but passing behavioral tests does not settle the question of conscious experience. The distinction between simulating intelligence and actually possessing subjective experience remains a central challenge for both philosophy and AI research.

### The frame problem

The frame problem, originally formulated in the context of symbolic AI, concerns how an AI system can efficiently determine which aspects of its knowledge remain unchanged when an action is performed. More broadly, it raises the question of how an intelligent agent can determine what is relevant in a given situation. Humans effortlessly filter irrelevant information, but formalizing this ability for AI systems has proven remarkably difficult. The frame problem has motivated significant work in knowledge representation, common-sense reasoning, and cognitive architecture design.

## Current state (2025 to 2026)

The AI field is advancing at an extraordinary pace, driven by massive investment, fierce competition among technology companies, and rapidly improving model capabilities. Stanford's 2025 AI Index reported that total corporate investment in AI reached $252.3 billion in 2024, including $33.9 billion in private funding directed specifically at generative AI, roughly eight times the 2022 level [36].

### Foundation models

The leading AI models as of early 2026 include:

| Model | Developer | Release | Notable capabilities |
|-------|-----------|---------|---------------------|
| [GPT-5](/wiki/gpt-5) | [OpenAI](/wiki/openai) | August 2025 | Multimodal reasoning, significantly reduced hallucination rates (up to 80% fewer factual errors than GPT-4) |
| Claude Sonnet 5 | [Anthropic](/wiki/anthropic) | February 2026 | 82.1% on SWE-Bench Verified (first model to break 80%), strong multi-step reasoning |
| Claude Opus 4.6 / Sonnet 4.6 | [Anthropic](/wiki/anthropic) | Early 2026 | 1-million-token context window |
| [Gemini](/wiki/gemini) 2.5 | [Google DeepMind](/wiki/google_deepmind) | Late 2025 | Strong reasoning capabilities, LMArena benchmark leader |
| [LLaMA](/wiki/llama) 4 | [Meta AI](/wiki/meta_ai) | 2025 | Open-weight model with strong tool orchestration |
| Grok | [xAI](/wiki/xai) | 2025 | Integrated with X (formerly Twitter) platform |

These models are increasingly multimodal, processing and generating text, images, audio, and video within single systems.

### Agentic AI

[Agentic AI](/wiki/ai_agent) emerged as a major trend in 2025 and 2026. Unlike traditional AI assistants that respond to individual prompts, agentic systems can autonomously perceive their environment, formulate multi-step plans, execute actions, use external tools, and adapt based on outcomes, all with minimal human intervention between steps [32].

McKinsey's State of AI 2025 report found that 88% of organizations were regularly using AI in at least one business function and that 62% were experimenting with or deploying [AI agents](/wiki/ai_agents) [32][34]. Gartner projected that 15% of day-to-day work decisions would be made autonomously through agentic AI by 2028, up from less than 1% in 2024 [32]. Anthropic's [Model Context Protocol](/wiki/model_context_protocol) (MCP), released in late 2024, became a widely adopted open standard for connecting AI agents to external tools and data sources, with OpenAI and Microsoft publicly embracing the protocol in 2025.

### Infrastructure investment

AI infrastructure spending has reached unprecedented levels. The major hyperscale cloud providers collectively planned to spend nearly $700 billion on data center projects in 2026, roughly double the approximately $365 billion spent in 2025. Individual company capital expenditure plans for 2026 include Amazon at approximately $200 billion, Google at $175 billion to $185 billion, and Meta at $115 billion to $135 billion [33].

[NVIDIA](/wiki/nvidia), the dominant supplier of AI training and inference chips, announced a partnership with OpenAI to deploy at least 10 gigawatts of NVIDIA systems. NVIDIA CEO [Jensen Huang](/wiki/jensen_huang) stated in March 2026 that he saw "at least $1 trillion" in computing demand through 2027 [33].

## Major AI research organizations

A number of companies and laboratories lead global AI research and development.

| Organization | Founded | Headquarters | Focus and notable contributions |
|-------------|---------|-------------|--------------------------------|
| [OpenAI](/wiki/openai) | 2015 | San Francisco, USA | GPT series, ChatGPT, DALL-E, Sora; surpassed $25 billion in annualized revenue by early 2026 |
| [Google DeepMind](/wiki/google_deepmind) | 2010 (DeepMind); merged with Google Brain in 2023 | London, UK | AlphaGo, AlphaFold, Gemini; two Nobel Prize-winning breakthroughs |
| [Anthropic](/wiki/anthropic) | 2021 | San Francisco, USA | Claude models, Constitutional AI, Model Context Protocol (MCP); focused on AI safety research |
| [Meta AI](/wiki/meta_ai) (FAIR) | 2013 | Menlo Park, USA | LLaMA open-weight models, PyTorch framework, foundational research in self-supervised learning |
| [Microsoft Research](/wiki/microsoft) | 1991 | Redmond, USA | Copilot, Azure AI; major investor in OpenAI; integrated AI across Office suite |
| [NVIDIA](/wiki/nvidia) | 1993 | Santa Clara, USA | H100, B200, Blackwell GPUs, CUDA; dominant hardware supplier for AI training and inference |
| [xAI](/wiki/xai) | 2023 | San Francisco, USA | Grok models; founded by Elon Musk |
| [Mistral AI](/wiki/mistral_ai) | 2023 | Paris, France | Mistral Large, open-weight models; leading European AI company |
| [Allen Institute for AI](/wiki/ai2) (AI2) | 2014 | Seattle, USA | OLMo open language models, Semantic Scholar; nonprofit research focus |
| [EleutherAI](/wiki/eleutherai) | 2020 | Distributed | GPT-NeoX, The Pile dataset; open-source, grassroots AI research collective |

Academic institutions also play a critical role, with Stanford HAI, MIT CSAIL, UC Berkeley's BAIR, Carnegie Mellon's School of Computer Science, the University of Toronto, and Mila (Quebec AI Institute, co-founded by Yoshua Bengio) among the most influential.

## See also

- [Machine learning](/wiki/machine_learning)
- [Deep learning](/wiki/deep_learning)
- [Large language model](/wiki/large_language_model)
- [Natural language processing](/wiki/natural_language_processing)
- [Computer vision](/wiki/computer_vision)
- [Artificial general intelligence](/wiki/artificial_general_intelligence)
- [AI alignment](/wiki/ai_alignment)
- [Generative AI](/wiki/generative_ai)
- [Transformer (machine learning)](/wiki/transformer)
- [Neural network](/wiki/neural_network)
- [AI regulation](/wiki/ai_regulation)

## References

[1] Russell, S. and Norvig, P. *Artificial Intelligence: A Modern Approach* (4th edition). Pearson, 2020.

[2] "History of artificial intelligence." Wikipedia. https://en.wikipedia.org/wiki/History_of_artificial_intelligence

[3] Turing, A.M. "Computing Machinery and Intelligence." *Mind*, vol. 59, no. 236, 1950, pp. 433-460.

[4] McCarthy, J., Minsky, M., Rochester, N., and Shannon, C. "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence." August 31, 1955. http://jmc.stanford.edu/articles/dartmouth/dartmouth.pdf

[5] Minsky, M. and Papert, S. *Perceptrons: An Introduction to Computational Geometry*. MIT Press, 1969.

[6] Lighthill, J. "Artificial Intelligence: A General Survey." Science Research Council, 1973.

[7] "Deep Blue versus Garry Kasparov." Wikipedia. https://en.wikipedia.org/wiki/Deep_Blue_versus_Garry_Kasparov

[8] "Watson, Jeopardy! champion." IBM History. https://www.ibm.com/history/watson-jeopardy

[9] Krizhevsky, A., Sutskever, I., and Hinton, G. "ImageNet Classification with Deep Convolutional Neural Networks." *Advances in Neural Information Processing Systems*, 2012.

[10] Silver, D., et al. "Mastering the game of Go with deep neural networks and tree search." *Nature*, vol. 529, 2016, pp. 484-489.

[11] Vaswani, A., et al. "Attention Is All You Need." *Advances in Neural Information Processing Systems*, 2017.

[12] Brown, T., et al. "Language Models are Few-Shot Learners." *Advances in Neural Information Processing Systems*, 2020.

[13] "Press release: The Nobel Prize in Chemistry 2024." NobelPrize.org. https://www.nobelprize.org/prizes/chemistry/2024/press-release/

[14] "ChatGPT launched three years ago today." TechCrunch, November 30, 2025. https://techcrunch.com/2025/11/30/chatgpt-launched-three-years-ago-today/

[15] "Press release: The Nobel Prize in Physics 2024." NobelPrize.org. https://www.nobelprize.org/prizes/physics/2024/press-release/

[16] "Navigating artificial general intelligence development: societal, technological, ethical, and brain-inspired pathways." PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC11897388/

[17] "Symbolic versus connectionist AI." AI Magazine, AAAI. https://ojs.aaai.org/aimagazine/index.php/aimagazine/article/download/15111/18883

[18] Bommasani, R., et al. "On the Opportunities and Risks of Foundation Models." Stanford HAI, 2021.

[19] "The State of Artificial Intelligence in 2025." Bay Tech Consulting. https://www.baytechconsulting.com/blog/the-state-of-artificial-intelligence-in-2025

[20] "The 2025 AI Index Report." Stanford HAI. https://hai.stanford.edu/ai-index/2025-ai-index-report

[21] Buolamwini, J. and Gebru, T. "Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification." *Proceedings of Machine Learning Research*, vol. 81, 2018.

[22] "Future of Jobs Report 2025." World Economic Forum, 2025.

[23] "Explained: Generative AI's environmental impact." MIT News, January 2025. https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117

[24] "AI Safety, Alignment, and Interpretability in 2026." Zylos Research, February 2026. https://zylos.ai/research/2026-02-09-ai-safety-alignment-interpretability

[25] "Statement on AI Risk." Center for AI Safety, May 2023. https://www.safe.ai/statement-on-ai-risk

[26] "AI Act." European Commission, Digital Strategy. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai

[27] "Highlights of the 2023 Executive Order on Artificial Intelligence for Congress." Congress.gov, Library of Congress. https://www.congress.gov/crs-product/R47843

[28] "Key Insights on President Trump's New AI Executive Order and Policy Regulatory Implications." Squire Patton Boggs. https://www.squirepattonboggs.com/insights/publications/key-insights-on-president-trumps-new-ai-executive-order-and-policy-regulatory-implications/

[29] "AI Risk Management Framework." NIST. https://www.nist.gov/itl/ai-risk-management-framework

[30] Searle, J. "Minds, Brains, and Programs." *Behavioral and Brain Sciences*, vol. 3, no. 3, 1980, pp. 417-424.

[31] Butlin, P., et al. "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness." *Trends in Cognitive Sciences*, 2023.

[32] "Agentic AI, explained." MIT Sloan. https://mitsloan.mit.edu/ideas-made-to-matter/agentic-ai-explained

[33] "Big Tech AI Infrastructure Spending 2026: The $700B Race." Tech Insider. https://tech-insider.org/big-tech-ai-infrastructure-spending-2026/

[34] "The state of AI in 2025: Agents, innovation, and transformation." McKinsey & Company, 2025. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

[35] "ChatGPT Officially Surpasses 800 Million Weekly Active Users." OpenAI DevDay announcement, October 2025. https://mlq.ai/news/chatgpt-officially-surpasses-800-million-weekly-active-users/

[36] "The 2025 AI Index Report: Top Takeaways." Stanford Institute for Human-Centered AI (HAI), April 2025. https://hai.stanford.edu/ai-index/2025-ai-index-report

[37] "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence (annotated full text)." Stanford University (jmc). https://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html