Artificial General Intelligence is a term used for a hypothetical or aspirational form of artificial intelligence possessing broad, flexible intellectual capabilities across many domains rather than being restricted to a narrow set of predefined tasks. There is no universally accepted technical definition or universally agreed operational test for Artificial General Intelligence, and researchers, companies and philosophers use the term in different ways.
Common interpretations involve the ability to learn new tasks, transfer knowledge between domains, reason about unfamiliar problems, adapt to changing circumstances, use previously acquired knowledge in new contexts and pursue goals across a broad range of environments. Some definitions emphasise performance comparable to humans across a wide spectrum of economically or intellectually relevant tasks, while others require stronger properties such as autonomous learning, robust world modelling or open-ended adaptation. Consequently, claims that a particular system constitutes Artificial General Intelligence depend partly on the definition being applied.
Artificial General Intelligence is commonly contrasted with narrow or specialised artificial intelligence. A narrow system may outperform humans on a particular benchmark or perform a highly complex task while remaining unable to generalise effectively outside its designed domain. Generality concerns breadth and transfer rather than simply maximum performance on one task. For example, a system that excels at image classification but cannot learn unrelated tasks without substantial redesign would ordinarily not satisfy broad interpretations of general intelligence.
Conversely, a system capable of learning diverse tasks from limited instruction, applying concepts across domains and adapting to novel situations would exhibit characteristics associated with generality. The distinction is complicated by modern foundation models, which can perform many tasks through a shared model and interface. Broad task coverage alone does not establish the stronger forms of general reasoning, autonomy or adaptability that some definitions of Artificial General Intelligence require.
The intellectual history behind Artificial General Intelligence extends through early artificial-intelligence research, cognitive science, philosophy of mind and computational theories of intelligence. Early artificial-intelligence researchers explored symbolic reasoning, theorem proving, planning, language understanding and general problem-solving systems. Some programmes attempted to encode knowledge explicitly through rules and logical representations, while others pursued learning from data. Neural approaches subsequently became increasingly important as statistical learning methods improved.
The modern rise of large pretrained models has renewed discussion because a single model can display capabilities across language, coding, reasoning-like tasks, image understanding and tool interaction. These developments have shifted the practical debate from whether machines can perform individual intelligent tasks to questions about the extent to which a common computational system can acquire and transfer abilities across diverse domains.
Several proposed properties are frequently associated with Artificial General Intelligence, although none is universally mandatory. General learning would involve acquiring capabilities without constructing a completely new system for every task. Transfer would involve applying knowledge learned in one setting to another relevant setting. Reasoning would involve deriving consequences, comparing alternatives and solving problems rather than merely reproducing memorised patterns. Planning would involve selecting sequences of actions to achieve goals. Adaptation would involve responding appropriately when conditions change.
Metacognitive capabilities could involve monitoring uncertainty or recognising limitations. Interaction with environments could involve perception, action and feedback. Long-term autonomy could involve maintaining objectives and executing extended plans. These properties are analytically distinct, so a system may exhibit some while lacking others. A meaningful assessment therefore requires specifying which capabilities and levels of reliability are being measured.
Evaluation of Artificial General Intelligence is unusually difficult because intelligence is multidimensional and benchmarks can become saturated or contaminated. A system may perform extremely well on familiar tests while failing under modest changes to instructions, environment or problem structure. Static benchmarks primarily measure performance on specified distributions and therefore cannot by themselves establish unrestricted generality. More informative evaluations may test unfamiliar tasks, transfer, adaptation, robustness, long-horizon planning, interaction with tools and ability to learn from limited examples.
There is also an important distinction between capability and autonomy: a system may be capable of solving a task when explicitly instructed but unable to identify appropriate objectives, acquire necessary resources or operate reliably over extended periods. Evaluation must therefore define the population of tasks, environmental assumptions, amount of available assistance and standards for success.
Artificial General Intelligence raises questions that extend beyond model architecture. A system could require substantial computational resources, external databases, tools, sensors and software infrastructure to perform broad tasks effectively. Generality might therefore emerge from the combination of a model with an environment rather than from a single parameter set. Questions about embodiment, persistent memory, causal understanding, world models and continual learning remain subjects of technical and theoretical investigation.
Another issue is reliability: broad capability is not equivalent to consistently correct behaviour. A system that can solve many kinds of problems but occasionally produces severe errors may be unsuitable for autonomous deployment in high-consequence environments. Robustness, uncertainty estimation, monitoring and mechanisms for human intervention would consequently remain important even if broad task competence were achieved.
The term is also frequently confused with human-level artificial intelligence, autonomous agents and large foundation models. These concepts overlap but are not identical. A foundation model is a technical model category defined by broad pretraining and reuse across tasks, whereas Artificial General Intelligence is a capability concept concerning breadth and flexibility. An autonomous agent is a system designed to perceive information, make decisions and execute actions, but an agent can be narrow or broad.
Human-level performance concerns comparison with human capabilities and may be specified differently from generality. A system could therefore have broad task coverage without matching humans in every cognitive domain, or could achieve superhuman performance in some areas while remaining limited in others. Precise terminology is necessary because claims about Artificial General Intelligence often depend on unstated assumptions about what counts as general intelligence.
The significance of Artificial General Intelligence lies partly in its role as a research objective and conceptual framework for examining the nature and limits of machine intelligence. Its possible development could have implications for scientific research, software engineering, economic production, education, automation, security and social institutions, but the timing and precise characteristics of such systems are uncertain. Discussions should distinguish documented capabilities of existing systems from predictions about future systems.
Technical progress in model scaling, multimodal learning, reasoning methods, tool use and autonomous operation can be measured independently without assuming that those developments establish Artificial General Intelligence. The concept is therefore most useful when treated as a carefully defined research question: what capabilities, generalisation properties, learning mechanisms, autonomy and reliability would a machine need to demonstrate before it could reasonably be described as possessing broad general intelligence?
