Artificial intelligence

Artificial intelligence (AI) is the capability of computational systems to perform tasks typically associated with human intelligence, such as learning, reasoning, problem-solving, perception, and language understanding. As an academic discipline, AI was founded at a workshop held at Dartmouth College in the summer of 1956, where the term was coined by John McCarthy. As of the mid-2020s, AI — in particular deep learning and large language models — has become one of the most transformative technologies in human history, reshaping industries from medicine to law and triggering intense debate about safety, governance, and the future of work.

Philosophical foundations

Questions about whether machines can think predate electronic computers. In 1637, René Descartes argued in Discourse on the Method that language and general reasoning distinguish humans from automata. In the 20th century, Alan Turing's 1950 paper "Computing Machinery and Intelligence" reframed the question operationally with the imitation game (now called the Turing test): if a human interrogator cannot reliably distinguish a machine's text responses from a human's, the machine may be said to exhibit intelligent behaviour.

Other influential philosophical positions include:

• The Chinese Room — John Searle's 1980 thought experiment arguing that symbol manipulation alone does not produce understanding, challenging the claims of "strong AI."
• Functionalism — the view (associated with Hilary Putnam and others) that mental states are defined by their functional role, not their physical substrate, providing philosophical support for the possibility of machine minds.
• The symbol grounding problem — Stevan Harnad's 1990 argument that formal symbols must be grounded in sensorimotor experience to carry meaning.

These debates remain unresolved and inform contemporary discussions about artificial general intelligence and consciousness in AI systems.

History

Early work (1943–1955)

The first mathematical model of an artificial neuron was proposed by Warren McCulloch and Walter Pitts in 1943. In 1950, Turing published his landmark paper. In 1951, Marvin Minsky and Dean Edmonds built SNARC, the first neural network computer. Claude Shannon and Turing independently explored chess-playing algorithms. By 1955, Allen Newell and Herbert Simon had created the Logic Theorist, often considered the first AI program, which proved theorems from Principia Mathematica.

The Dartmouth workshop and the golden age (1956–1974)

McCarthy, Minsky, Nathaniel Rochester, and Shannon organised the Dartmouth Summer Research Project on Artificial Intelligence in 1956, establishing AI as a field. The following years saw rapid progress:

• Newell and Simon's General Problem Solver (1957) — an early attempt at a general-purpose reasoning engine.
• ELIZA (1966) — Joseph Weizenbaum's natural language program that simulated a Rogerian psychotherapist, demonstrating how easily humans attribute understanding to machines.
• SHRDLU (1970) — Terry Winograd's natural language system for manipulating blocks in a simulated world.
• Perceptrons — Frank Rosenblatt's 1958 perceptron demonstrated simple pattern learning, but Minsky and Seymour Papert's 1969 book Perceptrons proved the single-layer perceptron could not learn XOR, contributing to reduced interest in neural networks.

Early AI was dominated by symbolic AI — the manipulation of human-readable symbols according to logical rules. Funding was generous, and predictions were optimistic: Simon predicted in 1965 that "machines will be capable, within twenty years, of doing any work a man can do."

First AI winter (1974–1980)

By the early 1970s, several foundational limitations had become apparent. James Lighthill's 1973 report to the British Science Research Council was harshly critical, and British government funding was cut dramatically. In the United States, DARPA reduced AI funding after combinatorial explosion made many problems intractable. The term "AI winter" was later coined to describe these periods of reduced funding and interest.

Expert systems and the boom (1980–1987)

Interest revived with expert systems — rule-based programs encoding domain knowledge. R1/XCON at Digital Equipment Corporation saved an estimated $40 million per year in computer configuration. The Japanese government launched the Fifth Generation Computer Systems project in 1982, spurring competitive investment worldwide. The AI industry grew from a few million dollars to over a billion dollars by 1985.

Second AI winter (1987–1993)

The expert systems market collapsed in the late 1980s as the systems proved brittle, expensive to maintain, and unable to learn. The desktop computer revolution made the specialised hardware (Lisp machines) obsolete. The Fifth Generation project failed to meet its goals. Funding again contracted.

Statistical turn and machine learning (1993–2011)

AI researchers increasingly adopted statistical and probabilistic methods — Bayesian networks, hidden Markov models, and support vector machines. These approaches, less ambitious in scope, produced reliable results in speech recognition, spam filtering, and recommendation systems. In 1997, IBM's Deep Blue defeated world chess champion Garry Kasparov. In 2011, IBM Watson won Jeopardy!. During this period, machine learning gradually displaced hand-engineered rule systems.

Deep learning revolution (2012–2017)

The modern era of AI began in earnest when Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton's AlexNet won the 2012 ImageNet Large Scale Visual Recognition Challenge by a wide margin, using a deep convolutional neural network trained on GPUs. This result demonstrated that deep networks with many layers, trained on large datasets with sufficient compute, could dramatically outperform traditional methods.

Key developments in this period:

• Generative adversarial networks (2014) — Ian Goodfellow's framework for training generative models through adversarial competition.
• Sequence-to-sequence models and attention mechanisms (2014–2015) — enabling breakthroughs in machine translation.
• Deep reinforcement learning — DeepMind's DQN playing Atari games (2013) and AlphaGo defeating Go world champion Lee Sedol (2016).
• Residual networks (ResNets, 2015) — enabling training of networks with hundreds of layers.

The transformer era (2017–present)

The publication of "Attention Is All You Need" by Vaswani et al. at Google in June 2017 introduced the transformer architecture, which replaced recurrence with self-attention. The transformer enabled massive parallelisation during training, leading to rapid scaling:

• BERT (2018) — Google's bidirectional pre-trained model, which set new records on 11 NLP benchmarks.
• GPT-3 (2020) — OpenAI's 175-billion-parameter autoregressive model, demonstrating strong few-shot learning.
• ChatGPT (November 2022) — brought large language models to mainstream public attention, reaching 100 million users in two months.
• GPT-4 (March 2023) — a multimodal model reportedly based on a mixture of experts architecture.
• Claude (2023–present) — Anthropic's family of models, trained using Constitutional AI.
• LLaMA (2023–2024) — Meta's open-weight models, catalysing the open-source AI movement.

By 2025, frontier AI models are trained on trillions of tokens of text and code using clusters of tens of thousands of GPUs, at costs exceeding $100 million per training run.

Core approaches

Symbolic AI

Also called Good Old-Fashioned AI (GOFAI), symbolic AI represents knowledge using human-readable symbols manipulated by logical rules. It was the dominant paradigm from the 1950s through the 1980s. Examples include expert systems, theorem provers, and planning systems. Its strengths are interpretability and the ability to encode known rules; its weaknesses are brittleness and the inability to handle uncertain or noisy real-world data ("the knowledge acquisition bottleneck").

Machine learning

Machine learning encompasses algorithms that improve through experience rather than being explicitly programmed. Major paradigms include:

• Supervised learning — learning from labelled input-output pairs (classification, regression).
• Unsupervised learning — finding structure in unlabelled data (clustering, dimensionality reduction).
• Reinforcement learning — learning to act in an environment to maximise a reward signal.
• Self-supervised learning — generating labels from the data itself (e.g., predicting the next token), now the dominant training method for large language models.

Deep learning

Deep learning uses artificial neural networks with many layers to learn hierarchical representations of data. Key architectures include convolutional neural networks (for vision), recurrent neural networks and long short-term memory networks (for sequences, largely superseded), and transformers (for language, vision, and multimodal tasks). Deep learning's success depends on three factors: large datasets, powerful hardware (GPUs and TPUs), and algorithmic advances like backpropagation, batch normalisation, residual connections, and the attention mechanism.

Neuro-symbolic AI

A growing research direction combining neural networks' pattern recognition with symbolic AI's logical reasoning. Examples include neural theorem provers and systems that use language models to generate logical programs.

Capabilities of modern AI systems

As of the mid-2020s, AI systems can:

• Understand and generate natural language — large language models produce fluent text, translate between languages, summarise documents, and answer questions, sometimes at or above human expert level on standardised tests.
• Perceive images and video — vision models classify objects, detect scenes, and generate images and video from text descriptions (via diffusion models and autoregressive vision models).
• Write and understand code — code-generation models can write, debug, and explain software in dozens of programming languages.
• Reason and plan — chain-of-thought prompting and search-augmented models exhibit multi-step reasoning, though with significant limitations in consistency and reliability.
• Predict scientific structures — AlphaFold predicted the 3D structure of nearly all known proteins; AI has accelerated drug discovery, materials science, and mathematics.

Applications

AI is deployed across virtually every industry:

• Healthcare — medical imaging diagnosis, drug discovery, clinical note summarisation, protein structure prediction.
• Finance — fraud detection, algorithmic trading, credit scoring, risk assessment.
• Transportation — autonomous vehicles (Waymo, Tesla, Cruise), route optimisation, traffic management.
• Science — automated experiment design, literature mining, simulation acceleration, theorem proving.
• Creative industries — image generation (Stable Diffusion, DALL-E, Midjourney), music composition, writing assistance.
• Software engineering — code completion (GitHub Copilot), automated testing, code review.
• Education — personalised tutoring, automated grading, adaptive learning platforms.
• Legal — contract analysis, case research, compliance monitoring.

Economic impact

AI is estimated by various analysts (McKinsey, Goldman Sachs, PwC) to contribute trillions of dollars to global GDP over the coming decade. Technology companies including Nvidia, Microsoft, Google, Meta, and Amazon have invested hundreds of billions of dollars in AI infrastructure. The rapid growth of AI has also raised concerns about labour displacement — the OECD and World Economic Forum have highlighted that AI could automate a significant fraction of existing jobs while creating new categories of work.

Safety, ethics, and governance

The rapid capability gains of AI systems have intensified concerns across several dimensions:

• Bias and fairness — AI systems can perpetuate and amplify biases present in training data, affecting hiring, lending, criminal justice, and other high-stakes decisions.
• Misinformation — generative AI can produce convincing but false text, images, and audio (deepfakes) at scale.
• Privacy — AI systems trained on internet data may memorise and reproduce personal information.
• Labour displacement — automation of cognitive tasks may displace white-collar workers in addition to traditional manufacturing roles.
• Concentration of power — the enormous cost of frontier AI training concentrates capability in a small number of well-funded organisations.
• Existential risk — researchers including Geoffrey Hinton, Yoshua Bengio, and many others signed a 2023 statement warning that "mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war."

Governance responses include the EU AI Act (passed 2024), US executive orders on AI safety, the UK and US AI Safety Institutes, China's AI regulations, and voluntary commitments by major AI labs.

Notable figures

• Alan Turing (1912–1954) — laid the theoretical foundations of computation and proposed the Turing test.
• John McCarthy (1927–2011) — coined the term "artificial intelligence" and invented Lisp.
• Marvin Minsky (1927–2016) — co-founder of the MIT AI Laboratory.
• Geoffrey Hinton (born 1947) — pioneer of backpropagation and deep learning; 2018 Turing Award and 2024 Nobel Prize in Physics.
• Yann LeCun (born 1960) — pioneer of convolutional neural networks; 2018 Turing Award; Chief AI Scientist at Meta.
• Yoshua Bengio (born 1964) — pioneer of deep learning and attention mechanisms; 2018 Turing Award.
• Fei-Fei Li — led the creation of ImageNet, which catalysed the deep learning revolution.
• Demis Hassabis (born 1976) — co-founder and CEO of DeepMind; 2024 Nobel Prize in Chemistry for AlphaFold.
• Ilya Sutskever (born 1986) — co-founder of OpenAI, co-author of AlexNet, later co-founded Safe Superintelligence Inc.
• Sam Altman (born 1985) — CEO of OpenAI.
• Dario Amodei (born 1983) — CEO of Anthropic, former VP of Research at OpenAI.

See also

• Machine learning
• Deep learning
• Large language model
• Transformer (machine learning)
• Artificial neural network
• AI safety
• AI alignment
• Artificial general intelligence
• Reinforcement learning from human feedback
• Natural language processing

References

• Turing, A. M. (1950). "Computing Machinery and Intelligence." Mind, 59(236), 433–460.
• McCarthy, J.; Minsky, M. L.; Rochester, N.; Shannon, C. E. (1955). "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence."
• Minsky, M.; Papert, S. (1969). Perceptrons: An Introduction to Computational Geometry. MIT Press.
• Lighthill, J. (1973). "Artificial Intelligence: A General Survey." Science Research Council.
• Searle, J. R. (1980). "Minds, Brains, and Programs." Behavioral and Brain Sciences, 3(3), 417–424.
• Krizhevsky, A.; Sutskever, I.; Hinton, G. E. (2012). "ImageNet Classification with Deep Convolutional Neural Networks." NIPS 2012.
• Vaswani, A. et al. (2017). "Attention Is All You Need." NIPS 2017.
• Brown, T. et al. (2020). "Language Models are Few-Shot Learners." NeurIPS 2020.
• Russell, S.; Norvig, P. (2020). Artificial Intelligence: A Modern Approach (4th ed.). Pearson.
• Bommasani, R. et al. (2021). "On the Opportunities and Risks of Foundation Models." Stanford CRFM.