AGI is not a single architecture but a research goal, so it has no one defined operating mechanism. Scientific discussion considers several paths toward AGI: scaling foundation models (the hypothesis that sufficiently large models and data yield emergent generality), agentic systems combining models with tools, memory and planning, world models and reinforcement learning, and neurosymbolic approaches combining learning with explicit reasoning. Progress is measured not by a single test but by multi-dimensional frameworks — DeepMind's "Levels of AGI" rates a system along two axes: depth (how well it performs tasks) and breadth (how many different tasks it covers), defining levels from "emerging" through "competent," "expert," "virtuoso" to "superhuman."
AGI addresses the limitation of narrow AI: systems achieving even superhuman results on single tasks fail to transfer competence to new problems or generalize beyond their training domain. The goal is a single system with general, flexible cognitive competence instead of many specialized models.
Alan Turing, in "Computing Machinery and Intelligence," proposes an operational test of machine intelligence, positing a machine with general intelligence comparable to a human.
The Dartmouth workshop establishes AI as a field of study, with the original, ambitious goal of reproducing general human intelligence in a machine.
John Searle introduces the distinction between "strong" and "weak" AI, providing philosophical context for later debates about general AI.
Mark Gubrud first uses the term "artificial general intelligence" in a discussion of the implications of fully automated production and operations.
Shane Legg and Ben Goertzel re-introduce and popularize the term AGI in the research community, distinguishing it from narrow AI.
The edited volume by Ben Goertzel and Cassio Pennachin consolidates AGI as a distinct research field and technical concept.
Nick Bostrom, in "Superintelligence: Paths, Dangers, Strategies" (Oxford University Press), conceptualizes artificial superintelligence and the risks of controlling it.
OpenAI's Charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work," giving the concept an influential operational definition.
Morris et al. propose a framework for classifying progress toward AGI by depth (performance) and breadth (generality), with levels from "emerging" to "superhuman."
