Artificial general intelligence, commonly abbreviated AGI, refers to a hypothetical form of Artificial intelligence capable of understanding, learning, and performing any intellectual task a human can, rather than excelling narrowly at a single domain. It is typically contrasted with "narrow AI," the specialized systems, from chess engines to image classifiers, that dominate deployed AI today. No system is widely agreed to have achieved AGI as of 2025, and the term's precise definition remains disputed among researchers, executives, and philosophers.
Definitions
There is no single accepted definition of AGI. Proposed criteria range from matching human performance across a broad battery of cognitive tasks, to possessing the ability to learn and transfer skills across domains without task-specific retraining, to economically meaningful thresholds such as a system that can perform most remote work tasks a human can. Some organizations, including OpenAI, have used contractual or economic definitions, tying AGI to a level of automated economic value creation, rather than purely technical ones, a choice critics say makes the term more a business milestone than a scientific benchmark.
History of the term
The phrase predates the current AI boom. Physicist Mark Gubrud used "artificial general intelligence" in a 1997 paper discussing nanotechnology and military risk, and the term was later adopted and popularized within AI research circles in the mid-2000s, notably through a 2007 edited volume titled "Artificial General Intelligence" associated with researcher Ben Goertzel. Its usage surged as dedicated organizations formed around the goal: OpenAI adopted AGI as its explicit mission when founded in 2015, and Google DeepMind, co-founded by Demis Hassabis as DeepMind in 2010 before its 2014 acquisition by Google, has similarly framed its mission around AGI since its founding.
Paths and predictions
Approaches proposed toward AGI include scaling existing Large language model and Foundation model architectures on more data and compute, the dominant industry bet since roughly 2020, combining language models with explicit planning and tool use as in AI agent systems, hybrid neuro-symbolic architectures that pair neural networks with logical reasoning, and approaches grounded in embodied interaction with the physical world. Surveys of AI researchers have shown wide disagreement on timelines, with median predictions for "high-level machine intelligence" shifting earlier across successive surveys conducted between 2016 and 2023, though individual estimates span from a few years to many decades.
Debate over proximity
Industry leaders have offered starkly different assessments. Sam Altman and other executives have described AGI as achievable within the current decade, while Hassabis has offered more measured timelines emphasizing remaining scientific gaps. The emergence of Reasoning model systems such as OpenAI o1 beginning in 2024, which allocate additional computation at inference time to multi-step problems, has been cited by some as evidence of accelerating progress toward general capability, and by others as an incremental improvement on narrow benchmark performance rather than genuine generality.
Criticism and skepticism
Researchers including Gary Marcus and François Chollet have argued that benchmark gains achieved by large language models often reflect memorization or interpolation over training data rather than robust generalization, pointing to persistent failures on novel reasoning problems. Chollet's ARC-AGI benchmark was explicitly designed to resist this critique by testing abstraction on tasks unlikely to appear in training data. Concerns about AGI are closely tied to broader debate over Superintelligence and Existential risk from AI, since many safety arguments assume that AGI, once reached, could rapidly self-improve beyond human oversight, a claim disputed by skeptics who see current architectures as fundamentally limited without further conceptual breakthroughs.