Loading field map
Artificial intelligence (AI) is the subfield of computer science concerned with building systems that exhibit intelligent behavior. Its central question is how to create computational agents—programs or machines—that can perceive their environment, reason about it, learn from experience, and act to achieve goals. The stakes are immense: AI promises to automate cognitive work, accelerate scientific discovery, and reshape economies, while also raising profound questions about human agency, fairness, safety, and the nature of intelligence itself.
AI is not a single problem but a cluster of interconnected challenges. The field asks: How can knowledge be represented so that a computer can use it? How can a system search through vast possibilities to find solutions or plans? How can it learn patterns from data without being explicitly programmed for every case? How can it understand and generate human language? How can it perceive the visual world? And how can it reason about uncertainty, causality, and the consequences of its own actions? These questions are not independent; progress in one area often depends on advances in others.
AI emerged as a distinct discipline in the mid-20th century, drawing on cybernetics, logic, and early computing. The term "artificial intelligence" was coined in 1956 at a workshop at Dartmouth College, where researchers gathered to explore the possibility of building machines that could think. The early decades were marked by optimism and a focus on symbolic reasoning: representing knowledge as logical statements and using deduction to solve problems. Programs that could prove mathematical theorems or play chess were seen as harbingers of general intelligence.
This symbolic approach, often called "good old-fashioned AI" (GOFAI), achieved notable successes but encountered fundamental limits. Real-world problems are messy, uncertain, and computationally intractable for exhaustive logical methods. By the 1970s and 1980s, the field experienced "AI winters"—periods of reduced funding and interest—as early promises failed to materialize. A parallel tradition, connectionism, which modeled intelligence as networks of simple units inspired by neurons, was largely marginalized during this period due to limited computational power and theoretical difficulties.
The late 1980s and 1990s saw a shift toward probabilistic reasoning and machine learning. Instead of hand-coding rules, researchers developed algorithms that could learn patterns from data. This statistical turn, combined with the growth of the internet and the availability of large datasets, led to dramatic progress in areas like speech recognition, computer vision, and natural language processing. The 2010s brought a resurgence of connectionism under the name "deep learning," enabled by powerful graphics processors and massive datasets. Deep neural networks achieved breakthrough performance on tasks from image classification to game playing, sparking a new wave of investment and public attention.
The field is organized less by a single sequence of paradigms than by enduring tensions and complementary methods. The most fundamental divide is between symbolic AI and connectionist AI, but these are not the only important traditions.
Symbolic AI treats intelligence as the manipulation of symbols according to explicit rules. Knowledge is represented in formal languages—logic, semantic networks, frames—and reasoning proceeds through search and inference. This approach addresses the problem of how to give a system a clear, interpretable model of the world. Its strengths are transparency (the system's reasoning can be inspected) and the ability to handle abstract concepts and long chains of deduction. Its limits are brittleness: symbolic systems struggle with ambiguity, noise, and knowledge that is hard to formalize. They also face the "knowledge acquisition bottleneck": encoding enough common-sense knowledge by hand is impractical. Symbolic AI remains influential in expert systems, automated planning, and formal verification, but it is rarely used alone for perception or learning from raw data.
Connectionist AI models intelligence as the emergent behavior of networks of simple processing units. Learning occurs by adjusting the strengths of connections between units based on examples. This approach addresses the problem of how to learn patterns from data without explicit rules. Its strengths are flexibility, robustness to noise, and the ability to handle high-dimensional data like images and audio. Deep learning, which uses many layers of units, has achieved remarkable results in pattern recognition, natural language processing, and game playing. Its limits include opacity (the learned representations are difficult to interpret), data hunger (it requires large labeled datasets), and difficulty with reasoning, planning, and causal inference. Connectionist systems also struggle with tasks that require systematic compositionality—combining known concepts in novel ways—though this is an active area of research.
A third major tradition treats intelligence as reasoning under uncertainty. Bayesian networks, Markov models, and probabilistic programming provide a framework for representing uncertain knowledge and updating beliefs based on evidence. This approach addresses the problem of how to make rational decisions when information is incomplete or noisy. It overlaps with both symbolic and connectionist methods: probabilistic graphical models combine logical structure with statistical inference, and modern deep learning often incorporates probabilistic elements. The strength of this tradition is its principled handling of uncertainty and its ability to combine prior knowledge with data. Its limits are computational: exact inference in complex probabilistic models is often intractable, requiring approximation.
Reinforcement learning addresses the problem of how an agent can learn to make sequences of decisions by interacting with an environment. The agent receives rewards or penalties for its actions and must learn a policy that maximizes cumulative reward. This approach is distinct from supervised learning (which learns from labeled examples) and unsupervised learning (which finds patterns in unlabeled data). Reinforcement learning has been central to achievements in game playing, robotics, and autonomous control. It can be combined with deep learning (deep reinforcement learning) to handle high-dimensional state spaces. Its limits include sample inefficiency (it often requires vast amounts of interaction) and difficulty with credit assignment (determining which actions led to a delayed reward).
No single approach has proven sufficient for general intelligence. Contemporary AI is increasingly hybrid: systems that combine symbolic reasoning with neural learning, or that integrate probabilistic inference with deep representations. For example, neural-symbolic systems attempt to give neural networks the ability to reason explicitly, while neuro-symbolic programming uses neural networks to learn components of a symbolic program. The field also draws on insights from cognitive science, neuroscience, and linguistics, though the relationship is often one of inspiration rather than direct modeling.
The current landscape is dominated by deep learning, which has become the default tool for perception, language, and generation tasks. Large language models (LLMs) and multimodal models—trained on vast text and image datasets—can generate coherent text, answer questions, write code, and create images. These systems are not intelligent in a general sense; they lack understanding, common sense, and reliable reasoning. They are pattern-matching engines that can produce plausible outputs but also hallucinate, exhibit biases, and fail on simple logical tasks.
The field is grappling with several enduring challenges. Interpretability remains a major concern: as models become more powerful, understanding why they make particular decisions becomes harder. Robustness is another: deep learning systems can be fooled by small, carefully chosen perturbations to inputs. Data efficiency is a practical limitation: current methods require enormous datasets, which are expensive to collect and may encode harmful biases. Safety and alignment have become central research areas: how to ensure that AI systems act in accordance with human values and goals, especially as they become more capable.
The relationship between AI and human intelligence is also contested. Some researchers pursue artificial general intelligence (AGI)—a system that can perform any intellectual task a human can—while others argue that intelligence is inherently situated, embodied, and social, and that the path to robust AI lies in building systems that interact with the physical world and with people over extended periods.
The field is no longer purely academic. AI is deployed in search engines, recommendation systems, medical diagnosis, autonomous vehicles, language translation, and creative tools. This deployment raises urgent questions about accountability, fairness, privacy, and the future of work. The technical and social dimensions of AI are now inseparable, and the field's central questions extend beyond algorithms to include the values and power structures embedded in the systems we build.