Artificial General Intelligence (AGI)

Artificial general intelligence (AGI) refers to a hypothetical class of artificial intelligence that can understand, learn, and apply knowledge across a wide range of intellectual tasks at a level comparable to, or exceeding, that of a human being. It is distinguished from the systems in widespread use today, which are forms of narrow or "weak" AI built to perform specific tasks such as language translation, image classification, or game playing. AGI, by contrast, would exhibit the breadth and flexibility of human cognition, transferring competence from one domain to another without being retrained for each new problem.

Definitions and disagreement

There is no single agreed-upon definition of AGI, and the lack of consensus is itself a central feature of the debate. Some researchers describe AGI as a system that matches or exceeds the cognitive versatility and proficiency of a well-educated adult across most economically valuable tasks. Others adopt a weaker standard, requiring only that a system perform "many but not all" of the tasks humans can. Because there is no universally accepted benchmark, there is also no agreed way to determine when AGI has been reached.

Two properties recur across most definitions. The first is generality, meaning the capacity to serve many purposes and accomplish different goals rather than a single fixed objective. The second is agency, meaning the ability to act in and interact with an environment to pursue those goals. A pragmatic formulation favoured by some researchers holds that AGI is achieved when it becomes difficult to devise problems that ordinary people can solve but that AI systems cannot.

Approaches and limitations

Much of the recent progress associated with the AGI debate has come from large language models trained on internet-scale text. These systems display surprising breadth, but a growing body of research argues that scaling models alone is not a sufficient path to general intelligence. Critics point to limits in genuine reasoning, planning, and out-of-distribution generalisation, where even advanced reasoning models can fail on problems that differ structurally from their training data. Proposed complements to pure scaling include neuro-symbolic methods, persistent memory, tool use, and architectures that combine learning with explicit world models.

The table below summarises common distinctions in the discussion.

| Term | Scope | Example | | --- | --- | --- | | Narrow AI (ANI) | Single task or domain | Spam filter, chess engine | | Artificial general intelligence | Human-level breadth | Hypothetical | | Superintelligence (ASI) | Beyond human ability | Hypothetical |

Timelines and safety

Expert predictions for the arrival of AGI vary widely, from those who expect early AGI-like systems to emerge within the next few years to those who regard it as decades away or fundamentally uncertain. Surveys of researchers and forecasters consistently show large disagreement, reflecting both definitional ambiguity and genuine scientific uncertainty.

Because a system with broad capability and autonomy could have far-reaching effects, AGI is closely tied to the field of AI safety and alignment, which studies how to ensure that highly capable systems pursue goals consistent with human intent. Concerns range from misuse and economic disruption to the difficulty of specifying and verifying objectives for systems more capable than their designers in particular domains.

References

1. IBM. (2024). What is Artificial General Intelligence (AGI)? IBM Think Topics.
2. METR. (2025). AGI: Definitions and Potential Impacts. Model Evaluation and Threat Research.
3. Morris, M. R., et al. (2024). Levels of AGI for Operationalizing Progress on the Path to AGI. Google DeepMind.
4. Council of Economic Advisers. (2026). Artificial Intelligence and the Great Divergence. The White House.