Artificial General Intelligence: Definitions, Levels, Open Questions
Artificial general intelligence (AGI) is a hypothetical machine intelligence that can match or exceed human cognitive abilities across any task. Researchers disagree on its exact definition and how to measure it, but several frameworks propose levels of performance and generality. Measuring progress toward AGI remains challenging because skill at specific tasks does not equal general intelligence.
.jpg){kind=link}
What Is Artificial General Intelligence
Artificial general intelligence is a hypothetical stage in machine learning where an AI system can match or exceed human cognitive abilities across any task, as described in IBM's explanation of AGI. It is a theoretical goal, not a current technology.
The term became popular through Ben Goertzel, who contrasted AGI with narrow AI, which only demonstrates intelligence in specialized domains. Most current AI systems are narrow, such as image recognizers or language models that cannot generalize beyond their training.
AGI is distinct from related concepts. Strong AI refers to a conscious mind, while artificial superintelligence exceeds human ability in virtually all tasks. The AWS overview of AGI explains that AGI is a theoretical pursuit to develop AI systems with autonomous self-control and the ability to learn new skills.
Key Definitions and Distinctions
A widely cited definition comes from François Chollet, who argues that intelligence is not skill at a task but skill-acquisition efficiency, or how quickly a system can learn new skills given prior knowledge and experience. This formal definition appears in his paper On the Measure of Intelligence on arXiv.
The ARC Prize Foundation's explanation of ARC-AGI adopts this view: AGI is a system that can match the learning efficiency of humans. It emphasizes generalization to novel problems rather than performance on a fixed set of tasks.
Another framework proposes levels of AGI based on performance and generality, along with autonomy. The Morris et al. levels paper suggests a common language to compare models and assess risks, but does not set a single threshold for AGI.
Proposed Levels of AGI
The levels framework from Morris et al. distinguishes depth (performance) and breadth (generality) of capabilities. It also considers autonomy, meaning how much human oversight a system requires when deployed. The paper argues that current systems fit at lower levels of this ontology.
There is no consensus on how many levels exist or what each level requires. Some proposals include levels from narrow to expert, with names such as exceptional for the highest tier, but these labels are not standardized across the field.
A different perspective from Google Cloud's article on AGI separates AGI from artificial narrow intelligence and artificial superintelligence, listing generalization and common sense as key traits of AGI. This highlights the need for transfer learning and world knowledge.
Measuring Intelligence and Open Questions
Chollet's paper argues that measuring skill at any given task is insufficient because unlimited data or priors can buy artificial skill, masking a system's generalization power. He proposes the Abstraction and Reasoning Corpus (ARC) as a benchmark to measure fluid intelligence using core knowledge priors.
The ARC Prize Foundation continues this work, describing ARC-AGI as a benchmark that is easy for humans but hard for AI, isolating the capacity for generalization. It restricts tasks to cognitive building blocks present at birth or acquired with minimal instruction.
Open questions remain: what exactly constitutes human-level intelligence, how to verify that a system truly generalizes, and how to compare AI with humans fairly. The levels paper stresses the need for future benchmarks that quantify behavior against these levels, and for careful selection of human-AI interaction paradigms to ensure safe deployment.
| Concept | Definition | Key characteristic |
|---|---|---|
| Narrow AI | Performs specific tasks within a defined domain | Cannot generalize to new domains |
| AGI | Matches or exceeds human cognitive abilities across any task | Generalization and self-teaching |
| Strong AI | Conscious mind, not merely a tool | Focus on consciousness, not performance |
| Artificial superintelligence | Exceeds human ability in virtually all tasks | Vastly superior capabilities |
| Weak AI | Tool for a conscious mind | Limited to programmed tasks |
Frequently asked questions
Is artificial general intelligence achieved?
No, AGI remains a theoretical goal. Current AI systems are narrow and cannot perform any intellectual task that a human can.
What is the difference between AGI and strong AI?
AGI focuses on broad human-level capability across tasks, while strong AI refers to a system with consciousness. They overlap but are distinct concepts.
How is intelligence measured in AGI research?
Researchers like François Chollet propose measuring skill-acquisition efficiency rather than task performance. The ARC-AGI benchmark uses core knowledge priors for fair comparison.
What are the levels of AGI proposed by researchers?
The Levels of AGI paper suggests levels based on performance (depth) and generality (breadth), plus autonomy, but the exact number of levels is not standardized.