What Is Artificial General Intelligence?

Artificial general intelligence, commonly abbreviated AGI, is the goal of building a machine whose competence transfers across arbitrary tasks and domains rather than being confined to the problems it was trained or engineered to solve. It is defined by contrast with narrow artificial intelligence, where a system trained for translation, protein structure prediction, or game play performs well within that task and cannot apply what it has learned elsewhere. The distinguishing property is generality itself: the capacity to acquire new skills efficiently, reason about unfamiliar situations, and carry knowledge from one setting into another without task-specific retraining.

The term became current in the early 2000s as a way of separating this ambition from the applied machine learning that dominated the field, though the underlying aim goes back to the founding proposals of artificial intelligence in the 1950s. AGI is a research objective rather than an existing technology, and there is no agreed test that would settle whether a given system had achieved it.

Definitions of Generality

Most working definitions treat AGI as matching the breadth of human cognitive performance across the range of tasks humans can perform, but that formulation leaves several questions open. Whether generality requires embodiment, whether it requires learning at human sample efficiency, and whether economic substitutability is the right criterion are all actively disputed. A recurring argument holds that current systems achieve breadth through exposure to enormous training corpora rather than through the compositional reasoning that would let them handle genuinely novel problems. One assessment of what stands between current models and artificial general intelligence argues that such systems still rely on people to formulate the problem, choose the architecture, and prepare the training data, and that passing a benchmark cannot show by what method the problem was solved. Persistent memory, causal reasoning, and continual learning without catastrophic forgetting are the capabilities most often named as unaddressed by scaling alone.

Research Approaches

Three lines of work dominate. Scaling large pretrained transformer models has produced the broadest capability gains to date and is the basis for the claim that generality may emerge from sufficient data, parameters, and compute. Agentic architectures wrap such models in planning loops, tool use, and external memory so that they can decompose long-horizon tasks and act on the results. Neurosymbolic and cognitive-architecture approaches instead argue that explicit structure, whether in the form of symbolic reasoning, learned world models, or biologically motivated organization, is needed for reliable generalization. Reinforcement learning contributes methods for goal-directed behavior and for aligning model outputs with human preferences.

Evaluation and Governance

Measuring progress toward generality is unresolved because existing benchmarks reward task performance and are vulnerable to contamination from training data. Proposals such as an operational testbed of foundational requirements for artificial general intelligence attempt to sidestep this by testing low-level cognitive prerequisites through reusable signal prediction tasks rather than language or perception scores. Evaluation also serves a safety function, since capability assessments now inform decisions about whether models are released, and the International AI Safety Report reviews the evidence base for those judgments across capability, risk, and mitigation. On the governance side, the NIST AI Risk Management Framework provides a voluntary structure for identifying and managing risks in AI systems, and it has become a common reference point for organizations deploying increasingly general models.

Applications

Research toward artificial general intelligence intersects with a range of fields, including:

• Foundation model development and large-scale distributed training
• Autonomous agents for software engineering, scientific discovery, and operations research
• Robotics and embodied learning
• Cognitive science and computational neuroscience
• AI safety, alignment, and interpretability research
• Standards, assurance, and policy work on advanced AI systems