Artificial General Intelligence

Artificial General Intelligence represents a hypothetical form of AI that would demonstrate a broad range of problem-solving, creativity, and adaptability comparable to human intelligence. Unlike narrow AI systems that excel at specific, well-defined tasks, AGI would possess the flexibility to understand, learn, and apply intelligence across diverse domains without explicit programming for each task.

The concept of general intelligence remains formally elusive. Legg and Hutter (2007) proposed an influential mathematical framework in their seminal work Universal Intelligence: A Definition of Machine Intelligence, defining machine intelligence as "an agent's ability to achieve goals in a wide range of environments." Their formalization uses Kolmogorov complexity to weight performance across all possible computable environments, providing a rigorous theoretical foundation for understanding what makes a system "generally intelligent".²

A system demonstrating AGI would theoretically perform cognitive tasks far beyond current AI systems. For instance, such a system could simultaneously engage in linguistic translation, creative composition in domains like music, and excel at novel problem-solving scenarios that have not been encountered before. This represents a qualitative leap from contemporary narrow AI systems, which are optimized for specific applications but lack the cross-domain adaptability inherent to human cognition.³

Researchers have identified several key distinctions necessary for understanding AGI's potential. Fluid intelligence—the flexible, abstract reasoning capacity to handle novel problems—would be central to AGI systems. Chollet (2019) proposed the Abstraction and Reasoning Corpus (ARC) as a framework for measuring such fluid intelligence, arguing that genuine general intelligence requires the ability to recognize abstract patterns and apply them to entirely novel situations.⁴

The path to AGI remains uncertain and contested within the research community. Current AI systems, despite demonstrating impressive narrow capabilities in domains such as image classification, natural language processing, and game-playing, fundamentally lack the flexibility and generalizability required for true AGI. Contemporary systems are brittle in unexpected ways—they fail on minor variations of tasks they have already mastered, suggesting their learned representations are task-specific rather than genuinely general.​⁵

The AI research community has identified multiple technical barriers to AGI development, including the lack of robust transfer learning mechanisms, limitations in few-shot learning capabilities, and insufficient theoretical understanding of how to combine specialized competencies into unified intelligent systems. Additionally, current neural network architectures, while powerful within their domains, have not demonstrated the compositional and causal reasoning capabilities that many theorists consider essential for general intelligence.⁶

The potential achievement of AGI raises profound questions that extend beyond technical considerations. A systematic review published in Journal of Experimental & Theoretical Artificial Intelligence identified multiple risk categories associated with AGI development, including the possibility of AGI systems removing themselves from human control, developing misaligned goals, and creating existential risks to humanity. The field currently lacks standardized terminology, specific risk modeling techniques, and adequate frameworks for managing AGI safety.​

Researchers have emphasized that AGI development raises critical ethical questions regarding value alignment—ensuring that an autonomous system's goals remain consistent with human values—and the need for governance frameworks that can adapt as AI capabilities advance.

Expert opinion on the timeline for achieving AGI varies considerably. Recent research surveying state-of-the-art large language models found estimates ranging from approximately 2030 to beyond 2040, with substantial uncertainty reflecting the fundamental challenges in predicting breakthrough moments in AI development. Most researchers agree that while significant progress has been made in narrow AI applications, the cognitive leap to AGI remains an open research challenge with no consensus on the approach most likely to succeed.⁷