Artificial Superintelligence refers to AI systems that substantially exceed human capability across all cognitive domains. The concept predates the modern deep-learning era — it was discussed philosophically by Nick Bostrom in the 2010s and by I. J. Good in the 1960s — and has gained new prominence as frontier models have approached or exceeded human performance on individual benchmarks.
ASI is distinguished from AGI primarily by degree. AGI refers to human-equivalent breadth; ASI refers to superhuman depth across that breadth. Debates about ASI typically center on whether its arrival would be gradual (via continuous AGI improvement) or discontinuous (via recursive self-improvement), and what governance structures would be required for its safe development.
No current AI system qualifies as ASI under any serious definition. Whether ASI is achievable, when, and under what conditions remain open questions at the frontier of AI research and policy.