Safety
Superintelligence
2014ResearchUpdated: 15 September 2026Published
Key
innovation
The notion of a hypothetical intelligence surpassing humans across all relevant domains, framing the horizon and terms of the debate on AI risk and control.
Category
Safety
Abstraction level
Paradigm
Operation level
SystemModel
Use cases
AI safety debateAI policy and regulationAlignment research
How it works
Superintelligence is not a specific architecture but a level of capability. The literature distinguishes possible paths to it (model scaling, recursive self-improvement, brain emulation) as well as takeover scenarios and risk-mitigation strategies (alignment, oversight, capability control).
Problem solved
Provides a conceptual frame for discussing the limits of AI progress and how to ensure safety and control over systems that surpass humans.
Evolution
Original paper · 2014 · Nick Bostrom
Superintelligence: Paths, Dangers, Strategies
Nick Bostrom
2014
Bostrom's "Superintelligence" book
Inflection pointNick Bostrom systematizes the concept of superintelligence, its paths and risks.