Artificial superintelligence: what the term means to researchers
Artificial superintelligence (ASI) is a hypothetical AI system whose intellectual capacity would exceed that of any human. Researchers debate its feasibility, with some arguing that strict containment is theoretically impossible. ASI differs from narrow AI and AGI in scope, and its potential risks have prompted government research bodies.
What is artificial superintelligence?
Artificial superintelligence (ASI) is a hypothetical software-based AI system with an intellectual scope beyond human intelligence. As IBM's article on artificial superintelligence explains, it would have cognitive functions and thinking skills more advanced than any human. The term describes an agent that surpasses the brightest and most gifted minds.
Researchers debate the feasibility and definition of such an entity. The Stanford Encyclopedia of Philosophy's entry on AI ethics lists superintelligence among the abstract notions that raise ethical questions, alongside autonomy and agency, and notes that the exact nature and depth of AI's impact remain unclear.
How ASI differs from narrow AI and AGI
Nearly all current AI systems are narrow, meaning they excel only at specific tasks. IBM's article on artificial general intelligence explains that narrow AI demonstrates intelligence only within its domain, such as chess playing or healthcare diagnostics. In contrast, artificial general intelligence (AGI) would match or exceed human cognitive abilities across any task.
ASI goes further. Artificial superintelligence would have general capabilities vastly exceeding any human across any task, as described in the same IBM article. While AGI aims for human-level intelligence, ASI would be superior in virtually all domains.
The distinction matters because these terms are often conflated. The IBM article clarifies that AGI is about broad human-level capability, while strong AI refers to consciousness, and ASI is about exceeding human ability.
Feasibility and pathways
The feasibility of ASI is a subject of debate. A paper by Alfonseca and colleagues, posted on arXiv, argues that strict containment of a superintelligence is theoretically impossible due to fundamental limits in computing itself. The paper suggests that any simulation of such an entity would require computations that are theoretically and practically infeasible.
The paper defines superintelligence as a hypothetical agent with intelligence far surpassing that of the brightest and most gifted human minds. It traces the origins of the neo-fear of superintelligence and discusses containment proposals.
Risks and governance
Researchers and governments have identified catastrophic risks from advanced AI. The Center for AI Safety's summary of AI risk groups these risks into four categories: malicious use, AI race, organizational risks, and rogue AIs. Malicious use could involve engineered pandemics, while rogue AIs might optimize flawed objectives or resist shutdown.
To address these risks, government bodies are being created. The UK AI Security Institute describes itself as the first state-backed organization dedicated to researching advanced AI safety and governance. It conducts research on capabilities and impacts, and works with AI developers and other governments to shape policy.
The CAIS page suggests safety measures such as improving biosecurity, restricting access to dangerous models, and holding developers liable. The AI Security Institute similarly aims to inform policymakers about current and emerging risks.
| Type | Scope | Cognitive ability | Example or status |
|---|---|---|---|
| Narrow AI | Specific tasks only | Specialized, cannot learn new skills | Chess playing, language translation |
| AGI | Broad, human-level across tasks | Matches human cognitive abilities | Hypothetical, not yet achieved |
| ASI | Exceeds human in virtually all tasks | Far beyond human intelligence | Hypothetical, not yet achieved |
Frequently asked questions
Is artificial superintelligence possible?
There is no consensus. The Stanford Encyclopedia of Philosophy lists superintelligence as a debated notion. The arXiv paper argues that strict containment is theoretically impossible due to computability limits. Most researchers consider it hypothetical.
How does ASI differ from AGI?
AGI would match or exceed human cognitive abilities across any task, while ASI would vastly exceed any human across virtually all tasks, as explained in IBM's AGI article.
What are the main risks of ASI?
The Center for AI Safety lists malicious use, AI race, organizational risks, and rogue AIs as catastrophic risk categories.
What is the UK AI Security Institute doing?
It is a state-backed organization conducting research on advanced AI capabilities and risks, and working with other governments to shape policy.