Artificial intelligence, commonly called AI, is the field of computing concerned with building systems capable of performing tasks that ordinarily require aspects of human or other forms of intelligent behaviour. Depending on the definition, these tasks can include perception, language processing, reasoning, planning, learning, prediction, decision-making, problem solving and generation of content. AI is not a single technology or algorithm.
It encompasses multiple approaches, including rule-based systems, search algorithms, statistical methods, machine learning, neural networks, reinforcement learning, knowledge representation and combinations of these techniques. In contemporary usage, AI often refers particularly to machine-learning systems that infer patterns from data, including large language models and generative systems, but the broader academic field predates modern machine learning and includes many non-learning approaches.
The concept of artificial intelligence has philosophical and technical roots predating electronic computers. Questions about whether machines can reason or exhibit intelligent behaviour were explored by philosophers and mathematicians long before modern computing. In the twentieth century, developments in mathematical logic, computation and electronic computers made the construction of machine-based reasoning systems increasingly practical.
The term artificial intelligence was prominently established as the name of a research field at the 1956 Dartmouth workshop, associated with researchers including John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon. Early AI research explored symbolic reasoning, theorem proving, game playing, planning and language processing. Progress occurred unevenly, with periods of enthusiasm followed by reduced funding and expectations known as AI winters.
Traditional AI approaches often represent knowledge explicitly through rules, logical statements, search spaces or structured symbolic representations. An expert system, for example, can use a collection of rules to infer conclusions from known facts. Search algorithms can explore possible sequences of actions to find solutions to planning or game-playing problems. These systems can be highly effective where the domain is sufficiently formalised and the rules are known, but they can struggle with ambiguous, noisy or highly variable real-world information. Machine learning changed the dominant approach in many applications by allowing systems to learn statistical relationships from examples rather than requiring programmers to encode every relevant rule manually. Modern AI frequently combines learned models with deterministic software, databases, search systems, tools and explicit constraints.
Modern AI includes several major application areas. Computer vision systems analyse images and video for tasks such as classification, detection, segmentation and recognition. Natural language processing systems analyse or generate human language. Speech systems convert speech to text or generate speech. Recommendation systems estimate which content or products may be relevant to users. Predictive models estimate outcomes from historical data. Generative AI systems produce text, images, audio, video or software code based on learned patterns and user inputs. Robotics combines perception, planning, control and physical interaction. These applications can rely on very different algorithms and data, so describing all AI systems as if they worked identically is technically inaccurate.
AI systems can be trained, programmed, or both. A rule-based system may behave according to explicitly authored logic without training on a dataset. A machine-learning model instead learns parameters from data through an optimisation process. Training data can consist of labelled examples, unlabelled data, demonstrations, feedback or combinations of these sources.
The resulting model does not necessarily store a simple database of the examples it saw; rather, learning changes internal parameters so that the model produces particular outputs when presented with particular inputs. The distinction between training and inference is important. Training is the process of fitting or adapting model parameters, while inference is the use of an already trained model to generate predictions, classifications, decisions or other outputs.
AI should not automatically be equated with human intelligence or consciousness. An AI system can perform a narrow task extremely effectively while lacking many abilities that humans possess across different contexts. Even highly capable models can make errors, rely on incomplete information, produce unsupported outputs or behave unpredictably outside their training and evaluation conditions.
Terms such as understanding, reasoning and intelligence can have different technical and philosophical meanings, and the observable performance of a system does not by itself establish that it possesses human-like subjective experience. Most deployed AI systems are designed for particular computational functions rather than being artificial persons. The field therefore concerns the engineering and study of intelligent-appearing computational behaviour, while questions about consciousness and general human-equivalent intelligence remain distinct research and philosophical issues.
AI systems have significant practical benefits but also important limitations and risks. They can automate repetitive analysis, identify patterns in large datasets, assist decision-making, improve accessibility, generate content and support scientific and engineering workflows. At the same time, AI can reproduce biases present in data, produce inaccurate outputs, expose sensitive information, create security vulnerabilities, consume substantial computational resources and make decisions that are difficult to explain. Generative systems can produce plausible but incorrect information, a phenomenon commonly described as hallucination. AI deployment therefore requires evaluation appropriate to the application, monitoring, security controls, data governance and human oversight where the consequences of errors are significant.
Artificial intelligence is best understood as a broad technical discipline rather than a synonym for one current technology. Machine learning is a major subfield of AI, and deep learning is a major approach within machine learning, while neural networks are a family of computational models heavily used in deep learning. Generative AI is an application category that can use several model architectures, including neural-network-based systems. AI systems may also combine machine learning with conventional algorithms, search, databases and externally supplied tools. This layered relationship explains why terms such as AI, machine learning, deep learning and neural networks are related but not interchangeable.
