A.i.

Artificial intelligence, conventionally abbreviated AI and occasionally punctuated A.I., is the branch of computer science concerned with computational systems that perform functions associated with intelligent behavior. These functions include deriving conclusions from represented information, identifying structure in sensory data, generating linguistic or visual material, and selecting actions in response to changing conditions. The field also examines the theoretical limits of such systems and the social institutions through which they are developed and deployed.

The term does not denote a single technology or an agreed computational definition of intelligence. It instead identifies a research domain whose boundaries have changed as particular capabilities became technically understood. Activities once treated as central demonstrations of machine intelligence, including arithmetic calculation and deterministic game search, subsequently became ordinary components of software engineering. This recurrent reclassification is known as the AI effect.

Conceptual foundations

The intellectual foundations of AI developed through work in mathematical logic, theories of computation, probability, and the study of biological nervous systems. Formal logic supplied methods for representing propositions and deriving valid conclusions from them, while probability theory provided a framework for reasoning under incomplete information. The theory of computation established which transformations could be performed by abstract machines and distinguished computational feasibility from logical possibility.

In 1950, Alan Turing analyzed the question of machine intelligence through an operational comparison between human and machine conversation. The resulting Turing test evaluates whether an interrogator can reliably distinguish generated responses from human responses under controlled conditions. It does not constitute a general definition of intelligence, consciousness, or understanding, although it became an influential model for behavioral evaluation.

Related developments in cybernetics treated adaptive behavior as the product of feedback between a system and its environment. This perspective connected early computing with control theory and with models of biological regulation. Artificial neural models provided another connection by representing computation as the transmission and adjustment of signals across networks of simplified units.

Formation of the field

Artificial intelligence became an identifiable academic field during the 1950s. The 1956 Dartmouth workshop introduced the term in a research proposal that treated learning and other aspects of intelligence as subjects for computational description. The workshop joined several previously separate traditions, including symbolic logic, automated search, neural modeling, and theories of information.

Working sessions involving Ray Solomonoff, Trenchard More, You Watanabe, and other participants examined how machines could form abstractions from experience and use internal representations to guide subsequent operations. Their discussions did not produce a unified architecture, but they established recurring questions concerning representation, learning, and the relation between general methods and task-specific knowledge. The workshop subsequently acquired greater historical coherence than it possessed as an actual research meeting, since its participants continued along substantially different technical paths.

Early programs demonstrated that computers could manipulate symbolic expressions rather than merely process numerical quantities. Allen Newell and Herbert A. Simon developed the Logic Theorist, which proved propositions from symbolic rules, and later produced the General Problem Solver as an investigation of domain-independent search. John McCarthy developed the Lisp programming language, whose treatment of programs and data as related symbolic structures influenced AI research for several decades.

These systems performed effectively within constrained formal environments. Their success did not transfer automatically to ordinary situations, where relevant knowledge is extensive, observations are incomplete, and the number of possible actions grows rapidly. The resulting disparity between demonstrations and general competence shaped subsequent cycles of institutional investment and contraction.

Symbolic systems and knowledge representation

The symbolic approach represents objects, relationships, and rules through explicit structures that a program can inspect and transform. Search algorithms then explore sequences of operations leading from an initial state to a specified goal. This framework supports mathematical proof, game playing, automated planning, and diagnosis when the relevant domain admits a sufficiently precise representation.

Symbolic systems encounter computational difficulty when the number of possible states expands exponentially. This phenomenon, commonly called combinatorial explosion, makes exhaustive search impractical even when every individual operation is simple. Heuristics reduce the explored space by assigning priority to states that are more closely associated with a solution, although their effectiveness depends on knowledge of the domain.

During the 1970s and 1980s, expert systems encoded specialized judgments as collections of rules. Systems such as MYCIN demonstrated that explicit knowledge bases could reproduce portions of professional decision-making within narrow settings. Their maintenance required substantial manual effort because new rules could conflict with existing rules, while tacit knowledge resisted conversion into discrete propositions. These limitations contributed to the commercial contraction known as the AI winter.

Statistical learning

Statistical machine learning shifted emphasis from manually encoded rules to models fitted from data. A learning algorithm adjusts internal parameters so that its outputs correspond more closely to observed examples or measured outcomes. The resulting model does not ordinarily contain a direct verbal statement of every regularity that influences its behavior.

Supervised learning uses examples paired with target outputs, while unsupervised learning derives structure from data without supplied labels. Reinforcement learning assigns consequences to sequences of actions and estimates strategies that increase cumulative reward. These categories describe the information available during training rather than the intellectual capacity of the resulting system.

Artificial neural networks regained prominence as larger datasets and specialized processors made multi-layer training computationally practical. Deep learning employs networks in which successive layers transform input representations into increasingly task-relevant internal patterns. Convolutional architectures became important in image processing, while recurrent architectures addressed sequential information before attention-based models displaced them in many linguistic applications.

Statistical success remains dependent on the relationship between training data and the environment of deployment. A model may rely on incidental correlations that disappear under changed conditions, producing failures known as distribution shifts. Performance measured on a fixed benchmark therefore describes behavior under that benchmark’s assumptions rather than establishing unrestricted competence.

Foundation and generative models

The transformer, introduced in 2017, uses attention mechanisms to represent relationships among elements of an input without processing them exclusively in serial order. Its architecture enabled the efficient training of models on large collections of text and later supported systems operating on images, audio, and other encoded media.

Large language models estimate conditional relationships among units of text and generate continuations from those estimates. Training at scale produces systems that perform translation, summarization, question answering, and program synthesis within a shared parameter structure. These capabilities arise from learned statistical regularities rather than from a database containing a separately stored answer for every prompt.

Generative systems also produce statements that are syntactically coherent but factually unsupported. This behavior, termed hallucination, follows from the distinction between predicting a plausible continuation and verifying a proposition against external evidence. Retrieval systems, formal tools, and structured databases can supply additional constraints, but their inclusion does not remove errors arising from ambiguous queries or defective source material.

The designation foundation model refers to a model trained on broad data and subsequently adapted to multiple applications. Such systems have altered the division between general-purpose infrastructure and application-specific software. They have not resolved the distinction between narrow task performance and artificial general intelligence, which denotes hypothetical systems able to transfer competence across domains with human-like generality.

Evaluation and interpretation

AI evaluation depends on the capability being measured and the conditions under which measurement occurs. Classification systems are assessed by comparing predicted categories with reference categories, while generative systems require criteria addressing factual consistency and relevance to a specified context. Interactive agents additionally require analysis of behavior across extended sequences, because an initially acceptable action may create later failures.

Benchmarks permit reproducible comparisons but influence research by turning selected tasks into optimization targets. Repeated exposure to benchmark data weakens the distinction between learning a general capability and adapting to a familiar test distribution. Independent evaluation therefore concerns not only numerical performance but also contamination of test material, sensitivity to prompt formulation, and stability under altered inputs.

The interpretation of learned models remains a separate research problem. Some internal features correspond to recognizable concepts, whereas other behavior is distributed across many parameters and cannot be reduced to a single rule. Explainable artificial intelligence studies methods for connecting outputs with influential inputs or internal computations, although an explanation that is understandable to a person is not necessarily a complete causal account of the model.

Social and institutional context

AI systems operate within organizations that determine their objectives, training material, access conditions, and acceptable error rates. Their effects consequently depend on institutional decisions as well as computational properties. Automated decisions can reproduce historical patterns contained in data, particularly when recorded outcomes reflect unequal treatment or incomplete measurement.

The collection and processing of training data also intersect with privacy, copyright, and information security. Generative models complicate conventional distinctions between memorization, transformation, and independent production because their outputs result from parameterized statistical processes while occasionally reproducing portions of training material. Legal systems address these issues through jurisdiction-specific rules rather than through a single technical definition of learning.

Research on AI safety examines failures arising from incorrectly specified objectives, unexpected environmental conditions, and interactions among autonomous components. Governance institutions separately regulate the contexts in which systems may be deployed and assign responsibility for resulting decisions. These activities treat AI as a class of engineered and administered systems rather than as an independent social actor.