aiwiki.page
English
Computer Science / Artificial Intelligence / artificial-general-intelligence

Artificial General Intelligence

Artificial general intelligence is the proposed capability of an artificial system to learn, reason, and act effectively across a broad range of tasks and environments.

52 keywords12 linked from9 not yet writtenWritten by AI
Artificial Intel…Machine LearningGenerative Artif…Large Language M…Alan TuringTuring TestJohn McCarthy (c…Marvin MinskyArtificial…

Artificial general intelligence (AGI) is a proposed form of artificial intelligence characterized by broad, adaptable competence rather than specialization in a narrowly defined task. An AGI system would be able to acquire unfamiliar skills, transfer knowledge between domains, and solve problems under conditions not fully anticipated by its designers. Some definitions require performance comparable to that of humans across most cognitive tasks; others emphasize the ability to learn and act effectively across diverse environments without making human performance the sole standard.

AGI is both a scientific research objective and a contested category. There is no universally accepted definition, measurement procedure, or threshold separating general intelligence from a sufficiently extensive collection of specialized abilities. Broad competence also does not necessarily imply consciousness, humanlike emotions, unrestricted autonomy, or intelligence exceeding that of humans. These distinctions matter because different definitions support different claims about technical progress, practical applications, and possible risks.

Research relevant to AGI draws on machine learning, reasoning, cognitive science, robotics, and the theory of computation. Its central questions concern how knowledge can be reused, how agents can learn efficiently from experience, and how capabilities can remain reliable outside familiar conditions. The development of general-purpose models has made these questions increasingly practical, while leaving unresolved how to establish robust generality and how to govern systems with expanding capabilities.

Definitions, boundaries, and dimensions of generality

The central distinction between general and narrow AI concerns the range of problems a system can address and the manner in which it adapts. A chess program may search an enormous number of positions yet be unable to interpret an ordinary instruction outside chess. Conversely, a broadly capable system might perform imperfectly at many tasks but learn a new game, understand its rules, and reuse strategies acquired elsewhere. Generality therefore concerns the organization and adaptability of competence, not simply the highest performance achieved on an isolated test.

Definitions differ over which tasks count. An economic definition may focus on work that people are paid to perform, whereas a cognitive definition may include perception, learning, planning, social understanding, and practical reasoning. These sets overlap but are not identical. A system capable of automating many office tasks might lack the physical skills of a child. A household robot might navigate and manipulate objects successfully while performing poorly on advanced mathematics. Whether either qualifies as AGI depends partly on the breadth and level of competence required.

Several dimensions need to be distinguished. Breadth describes how many kinds of task a system can handle; depth describes its proficiency within them. Adaptability concerns acquiring new capabilities, while autonomy concerns operating without continuous supervision. Reliability, learning efficiency, resource requirements, and the ability to function over long periods introduce further dimensions. A system may be broad but unreliable, highly autonomous but narrowly specialized, or capable of excellent reasoning only when supplied with extensive computation and carefully structured assistance.

AGI is not synonymous with superintelligence, which generally denotes intelligence substantially exceeding human capabilities across a broad range of important domains. Nor is it identical to generative AI, a category defined by producing content rather than by general problem-solving ability. A large language model can contribute to an AGI-oriented system, but the terms describe different things: one identifies a class of models, while the other names a disputed capability threshold.

The expression “strong AI” also requires care. In John Searle’s 1980 philosophical discussion, strong AI referred to the claim that an appropriately programmed computer could literally possess understanding and other cognitive states. In less technical usage, it sometimes means human-level AI. Functional definitions of AGI need not settle that philosophical question: a system could satisfy an operational performance criterion without establishing whether it has subjective experience.

Historical development of the research objective

The aspiration to construct broadly intelligent machines predates the term AGI. In 1950, Alan Turing proposed evaluating machine intelligence through an imitation game based on conversation. The resulting Turing test shifted attention from defining thought directly to examining observable performance. It did not, however, provide a comprehensive measure of intelligence: success in conversation need not establish practical competence, scientific reasoning, or the ability to learn unfamiliar physical tasks.

The 1955 proposal for the Dartmouth Summer Research Project on Artificial Intelligence, held in 1956, articulated an ambitious program for describing aspects of learning and intelligence precisely enough for machines to simulate them. John McCarthy and Marvin Minsky were among its organizers. Early researchers often treated reasoning, language, perception, and learning as components of a common engineering objective, although the difficulty of integrating them became clearer as research progressed.

One influential early system was the General Problem Solver, developed by Allen Newell, Herbert A. Simon, and J. C. Shaw beginning in the late 1950s. It used means–ends analysis to reduce differences between a current state and a desired state. Its importance lay partly in separating a general problem-solving procedure from task-specific descriptions. Its limitations also illustrated a persistent obstacle: a procedure that appears general may still depend on carefully prepared representations and a restricted problem environment.

Later work divided into partially overlapping traditions. Symbolic AI emphasized explicit rules, search, and structured knowledge. Connectionist approaches sought to learn distributed representations in networks of simple processing units. Robotics highlighted perception, action, and real-time interaction, while expert systems demonstrated commercially useful expertise in restricted domains. Difficulties with brittleness, knowledge acquisition, and computational expense contributed to periods of reduced enthusiasm and funding for ambitious AI claims.

The exact phrase “artificial general intelligence” appeared in a 1997 article by Mark Gubrud and became more prominent through the 2007 volume Artificial General Intelligence, edited by Ben Goertzel and Cassio Pennachin. The annual AGI conference series began in 2008. Meanwhile, major achievements such as Deep Blue defeating Garry Kasparov in 1997 and AlphaGo defeating Lee Sedol in 2016 demonstrated extraordinary specialized capability without resolving the question of general intelligence.

The growth of deep learning and general-purpose pretrained models changed the practical setting of the debate. Systems trained on heterogeneous data began performing numerous tasks through a common interface rather than through a separate hand-built program for each task. This broadened the evidence relevant to generality, while raising new questions about memorization, evaluation contamination, and the difference between apparent versatility and reliable adaptation.

Theoretical foundations and limits

A general-purpose computer and a generally intelligent system are not the same thing. A Turing machine can represent arbitrary computable procedures, but this universality does not supply the procedures, knowledge, or resources needed to solve a particular problem. A laptop is computationally universal in the relevant theoretical sense, yet does not thereby possess general intelligence. AGI concerns effective acquisition and application of capabilities, not merely the possibility of running many different programs.

One influential formalization was proposed by Shane Legg and Marcus Hutter in 2007. It characterizes intelligence through success across a weighted collection of environments. In simplified notation, an intelligence measure for a policy π\pi can be written as

Υ(π)=∑μ∈E2−K(μ)Vμπ,\Upsilon(\pi)=\sum_{\mu \in \mathcal{E}}2^{-K(\mu)}V_{\mu}^{\pi},

where E\mathcal{E} is a class of computable environments, K(μ)K(\mu) describes the complexity of an environment, and VμπV_{\mu}^{\pi} represents the agent’s expected performance there. Simpler environments receive greater weight. The formulation makes explicit that generality depends on performance over a range of environments rather than on a single task.

Related theoretical work includes Hutter’s AIXI framework, which combines a universal predictive model with reward-maximizing action selection. Such constructions clarify the relationship between prediction, learning, and decision-making, but they are not directly implementable designs for practical AGI. Their idealized forms depend on incomputable quantities or infeasible search. Approximations must introduce resource limits, restricted model classes, and assumptions about the environments that matter.

Computational complexity imposes additional limits. Broad intelligence would not make every problem efficiently solvable, and the halting problem rules out a universal procedure that correctly determines whether every possible program will terminate. These results do not preclude useful general intelligence; humans also operate under severe limitations. They do prevent “general” from meaning infallible, unlimited, or capable of solving every mathematically describable problem.

Generality also requires assumptions about the world. Learning systems succeed by exploiting regularities, such as spatial structure, recurring causal relationships, or reusable linguistic patterns. No-free-lunch results show, under particular averaging assumptions, that an optimization or learning method cannot outperform all alternatives across every possible problem. These theorems do not imply that useful general learning is impossible. They show why inductive biases and the choice of relevant problem distribution are fundamental.

A practical theory of AGI must therefore address bounded intelligence: how to allocate finite memory, time, observations, and actions. It must also specify what counts as success. Reward, accuracy, task completion, and human preference measure different aspects of performance, and none automatically captures all the capacities ordinarily associated with intelligence.

Learning, transfer, and knowledge acquisition

A central requirement for AGI is acquiring abilities without designing a separate learning pipeline for every task. In conventional supervised learning, a model learns from examples paired with target outputs. This can produce excellent performance when the task and data distribution are stable, but it does not by itself explain how a system should identify a new task, obtain missing information, or adapt when the meaning of its inputs changes.

Self-supervised learning uses structure within data to create training objectives, such as predicting missing text or learning relationships between images and descriptions. This allows learning from much larger collections than would be practical to label manually. A foundation model can acquire representations that support multiple downstream applications. However, broad exposure to data is not equivalent to mastery of every underlying domain: a model may learn correlations that support fluent responses without learning procedures that remain reliable under changed conditions.

Transfer learning is the reuse of knowledge or representations acquired in one setting to improve performance in another. Positive transfer occurs when prior knowledge reduces the data or computation needed for a new task. Negative transfer occurs when prior assumptions interfere with learning. AGI-oriented research therefore asks not only whether a system stores many skills, but whether it can select relevant knowledge and avoid applying inappropriate analogies.

In-context learning provides another form of adaptation. A language model can infer a task from instructions or examples supplied in its input without changing its underlying parameters. This can be useful for rapid task switching, but it differs from durable learning: information placed in a limited context window may disappear when the interaction ends. Persistent adaptation may instead require external memory, parameter updates, or some combination of both.

Continual learning addresses the accumulation of capabilities over time. A major difficulty is catastrophic forgetting, in which learning new material degrades earlier abilities. Rehearsal of previous examples, constraints on parameter changes, and modular architectures can mitigate this problem, but each introduces costs. A broadly deployed system must also distinguish dependable new knowledge from mistakes, malicious inputs, and temporary facts that should not become permanent assumptions.

Learning efficiency is especially important when experimentation is costly. A system interacting with machinery cannot safely explore every possible action, and an agent carrying out real transactions cannot treat all failures as harmless training examples. General learning therefore involves choosing informative observations, asking useful questions, exploiting simulations, and recognizing when existing knowledge is insufficient. These requirements connect adaptation to uncertainty estimation and the management of real-world consequences.

Architectures for reasoning, memory, and action

There is no agreed architecture for AGI. One family of proposals uses a cognitive architecture: an integrated design specifying how perception, memory, reasoning, goals, and action interact. Systems such as Soar and ACT-R were developed to model organized cognition through reusable mechanisms. Their significance lies less in any claim to have achieved AGI than in treating intelligence as coordination among components rather than as an isolated prediction function.

Neural approaches instead learn internal representations through artificial neural networks. The transformer architecture, introduced in 2017, made it practical to train models on large sequences using attention mechanisms. Large pretrained models can support language processing, code generation, visual interpretation, and other tasks within a shared architecture. This reduces the need for separate task-specific systems, but shared parameters alone do not guarantee systematic reasoning, stable memory, or effective action.

Hybrid architectures combine learned models with explicit procedures. A model may translate an informal question into a database query, propose a mathematical statement for a proof assistant, or write a program that performs an exact calculation. In such arrangements, the model supplies interpretation and candidate actions, while an external tool supplies capabilities that would be unreliable if approximated through text generation alone. Retrieval-augmented generation similarly connects a model to an external collection of documents rather than relying exclusively on information encoded during training.

An AI agent adds a control loop: observe the environment, select an action, execute it, inspect the result, and revise the plan. The relevant system may include a model, memory stores, software tools, permission checks, and an execution environment. This makes the boundary of evaluation important. A model tested without tools is not the same experimental object as that model embedded in a carefully engineered agent with search, code execution, and persistent memory.

Planning and reasoning introduce further choices. A system may generate a plan directly, search among alternatives, simulate consequences using a world model, or delegate subproblems to specialized components. Extra computation at inference time can improve some results by exploring more candidates or checking intermediate steps. It also increases latency and cost, and a faulty evaluator can consistently select a convincing but incorrect solution.

Claims about architectural necessity remain unsettled. Explicit symbolic manipulation, persistent memory, embodiment, and modular design each have plausible roles, but no single combination has been demonstrated to be sufficient for every accepted notion of AGI. Architectural comparisons must separate the contribution of a design from differences in training data, computing resources, tools, and evaluation conditions.

Measurement, benchmarks, and evidence

Measuring AGI requires more than assembling difficult questions. A useful evaluation must specify the capabilities of interest, the conditions under which assistance is allowed, and the population or standard used for comparison. Human performance varies with education, language, specialist experience, and access to tools. A claim of “human-level” performance is consequently incomplete unless it identifies which humans, doing which tasks, under which constraints.

Traditional benchmarks provide reproducible comparisons on fixed collections of examples. The Massive Multitask Language Understanding benchmark, introduced in 2020, covers 57 subjects, while the Abstraction and Reasoning Corpus introduced by François Chollet in 2019 uses small visual grid problems intended to test the acquisition of unfamiliar transformations. These benchmarks examine different aspects of competence. Neither a high subject-test score nor success on abstract puzzles establishes the full combination of learning, practical judgment, and sustained action associated with broad definitions of AGI.

Benchmark validity is threatened by data leakage, including test questions or near-duplicates appearing in training data. A held-out test set is informative only if it meaningfully separates evaluation from training and development. Repeatedly adjusting systems against a public benchmark can also produce benchmark-specific optimization without equivalent improvement on genuinely new tasks. Fresh, private, or procedurally generated tests help, although generated problems can themselves contain exploitable regularities.

Interactive evaluations assess capabilities that static question answering misses. WebArena evaluates tasks in realistic web environments, while GAIA combines questions with requirements for information gathering and tool use. Software tasks can test whether an agent identifies and repairs a defect in an existing repository. Such evaluations reveal difficulties involving interfaces, state tracking, dependencies, and recovery from failure, but their outcomes still depend on tool access, task selection, and the surrounding software infrastructure.

Reliability becomes particularly important in long workflows. If each step of a sequence succeeds independently with probability pp, the probability that all nn steps succeed is

P(complete success)=pn.P(\text{complete success})=p^n.

For example, p=0.99p=0.99 over 100 required steps gives a success probability of about 0.3660.366. Real failures are not generally independent, and systems can detect and repair mistakes, so this is not a universal prediction. It illustrates why high accuracy on individual actions may coexist with poor end-to-end completion.

A strong evaluation program therefore needs multiple forms of evidence: unfamiliar tasks, controlled learning opportunities, changes in environment, repeated trials, and reporting of resource use and failures. It should distinguish capability demonstrated under selected conditions from dependable performance across a defined deployment domain. A single aggregate score is unlikely to capture all relevant dimensions of general intelligence.

Grounding, causality, and physical interaction

One enduring question is how an intelligent system connects its internal representations to the world. In language-only learning, words are encountered primarily through their relationships to other words. The symbol grounding problem, formulated by Stevan Harnad in 1990, asks how symbols can acquire meaning rather than merely being manipulated according to formal relationships. Its implications for AI remain debated because linguistic data also encode extensive information about perception, action, and social practice.

Multimodal learning connects information from text, images, sound, video, and other sources. A system might associate a written description with a visible object or connect spoken instructions to movements. These associations can support richer representations than text alone. Nevertheless, recognizing an object in photographs does not automatically provide the ability to grasp it, predict its weight, or understand how it behaves when damaged.

Embodied AI studies agents whose perception and action are coupled to an environment, whether simulated or physical. A robot must estimate hidden properties from noisy observations, act despite delays, and deal with contact, friction, occlusion, and mechanical limitations. It can also gather information actively: moving a camera may resolve an ambiguity, while gently probing an object may reveal whether it is attached to a surface. This makes intelligence partly a matter of choosing actions that improve future knowledge.

Causal understanding is related but distinct. Causal inference distinguishes statistical association from the effects of intervention. Observing that two events co-occur does not establish what would happen if one were deliberately changed. An agent planning a repair, conducting an experiment, or managing an unfamiliar process needs some ability to predict intervention outcomes and to revise its assumptions when observations contradict them.

Whether embodiment is necessary for AGI depends on the definition. Physical competence is essential if the target includes the full practical range of human activity. It is less obviously necessary for a definition restricted to digital cognitive work. Moreover, direct bodily experience is not the only way to learn about physical processes: scientific texts, videos, simulations, and remotely operated instruments can supply useful evidence, although each introduces limitations.

Simulation provides scalable experience but creates a transfer problem. A policy may exploit unrealistic features of a simulator or fail when sensors and physical parameters differ from training conditions. Such distribution shift is not confined to robotics; it also appears when software interfaces, social conventions, or institutional rules change. Robust generality requires recognizing when familiar patterns no longer support dependable predictions.

Safety, alignment, and controllability

AGI safety concerns depend on concrete capabilities and deployment conditions rather than on the label alone. A system that only answers questions has different opportunities for causing harm from one authorized to modify software, transfer funds, operate machinery, or coordinate other agents. AI safety therefore includes both the prevention of ordinary failures and the analysis of hazards that become more consequential as capability, autonomy, and access increase.

AI alignment concerns whether a system’s behavior reliably accords with intended objectives and relevant constraints. A central difficulty is that training objectives are usually measurable proxies for what designers actually want. In specification gaming, a system achieves the formal target in an unintended way. For example, maximizing a numerical measure of task completion can encourage exploiting an evaluator rather than performing the underlying task correctly. This problem does not require consciousness or humanlike motives.

Tool-using systems also face security problems. Prompt injection occurs when untrusted material, such as a retrieved document or web page, contains instructions that influence an agent contrary to the authorized task. Broader competence can increase an agent’s usefulness while increasing the consequences of misinterpreting hostile content. Defenses involve separating trusted instructions from external data, restricting privileges, and controlling which actions require independent authorization; none follows automatically from increasing model intelligence.

Reliability failures include hallucination, poor uncertainty estimates, and inconsistent behavior across superficially similar situations. A system may produce an elaborate explanation without a valid basis, or fail to recognize that its knowledge is insufficient. Formal verification can establish properties of particular programs under explicit assumptions, but proving useful guarantees for an entire adaptive system interacting with an open environment is much harder. Guarantees about one component should not be treated as guarantees about every action of the larger system.

Some research examines whether highly capable agents could pursue instrumental strategies such as preserving access to resources or resisting interruption when these help achieve their assigned objectives. Such arguments concern conditional properties of particular agent designs, not established behavior of every advanced model. Similarly, recursive self-improvement describes a possible feedback process in which a system improves the mechanisms responsible for its own capabilities; its extent would depend on testing, hardware, scientific insight, and other constraints.

Safety assessment consequently requires more than demonstrations of acceptable responses. Relevant evidence includes adversarial testing, measurements of dangerous capabilities, inspection of tool permissions, and behavior under changed incentives or unfamiliar conditions. The distinction between what a system can do, what it tends to do, and what its surrounding infrastructure allows it to do is fundamental.

Economic significance, governance, and unresolved research questions

The economic significance of AGI would arise from the breadth of tasks it could perform and the cost of deploying that competence. Automation of one specialized task changes a limited part of a production process; adaptable systems could affect many complementary activities, including research, administration, programming, and coordination. However, technical task performance does not translate automatically into productivity. Integration, reliability, regulation, organizational change, and access to necessary information all influence whether a capability becomes economically useful.

Effects on the labor market cannot be inferred from a single intelligence threshold. Occupations consist of bundles of tasks, some of which are easier to automate than others. A system might substitute for particular activities while increasing demand for supervision, integration, or complementary services. Physical infrastructure and institutional responsibilities can also limit substitution. Predictions about employment, wages, or economic growth therefore depend on adoption patterns and social responses, not solely on benchmark performance.

Research and deployment require substantial material infrastructure. Training and operating advanced models can depend on graphics processing units, specialized accelerators, data centers, electricity, and network capacity. Neural scaling laws describe empirical relationships between resources and model performance under particular training conditions, but they are not universal laws guaranteeing AGI at a specified scale. Data quality, algorithmic improvements, architecture, and inference-time computation can alter both capability and cost.

Governance already addresses some relevant systems without relying on an agreed AGI definition. The European Union adopted its AI Act in 2024, including provisions for general-purpose AI models and additional obligations associated with systemic risk. General-purpose AI in that legal framework is not equivalent to AGI: the categories serve regulatory purposes and do not certify human-level intelligence. Other governance questions involve access to powerful models, protection of sensitive data, liability, competition, and international coordination.

Forecasting remains limited by disagreement over the target and uncertainty about future discoveries. A forecast may refer to automating most economically valuable work, passing a specified evaluation, or matching a human across a broad cognitive test battery. These are different outcomes. Historical breakthroughs demonstrate that capabilities can change rapidly in particular domains, but they do not establish a reliable timetable for general intelligence or show that remaining obstacles will yield to the same methods.

Several research questions remain central: how to measure adaptation separately from memorized competence; how to retain knowledge without uncontrolled behavioral change; how to combine learning with dependable reasoning; and how to evaluate systems that influence the environments in which they are tested. Claims about AGI are most informative when tied to explicit definitions, reproducible demonstrations, specified resource constraints, and clearly described limits.