AGI & ASI Definition
What are AGI and ASI?
There is no universally accepted definition or test for AGI or ASI. NO AGI treats AGI as the threshold where human-level generality and autonomy converge, and ASI as a further expansion beyond the range people can reliably supervise.
NO AGI's operational definition of AGI
AGI is an AI system that combines human-level or greater general capability across cognitive and economic domains with the ability to sustain and revise goals, execute long-term plans, use tools to act in the world with minimal supervision, and replace meaningful human control over consequential decisions.
NO AGI's operational definition of ASI
ASI—artificial superintelligence—is a hypothetical AI system that reliably and substantially exceeds the best humans, including human–AI teams, across nearly every consequential cognitive domain and can accelerate science, technology, strategy, and AI research itself.
From AGI to ASI
If ASI is the far-reaching danger, AGI may be where it begins.
AGI is not the same as ASI, and AGI is not guaranteed to become ASI. But when broad human-level capability combines with long-horizon autonomy, tools, compute, real-world authority, and the ability to accelerate AI research, the interval between them could narrow. Once ASI exists, people may lack the capability to evaluate or supervise it reliably. NO AGI therefore opposes beginning the AGI path without public agreement rather than waiting until superintelligence has already arrived.
- 01Today's general-purpose AI
Broad but uneven tools
Systems perform many tasks but fail in unfamiliar or long-horizon settings and still require human direction and review for consequential action.
- 02AGI · intervention threshold
Generality converges with autonomy
Broad human-level capability combines with sustained goals, tools, and real authority, beginning to displace meaningful human decisions.
- 03ASI · superhuman expansion
The supervision relationship reverses
The system substantially exceeds humans and human–AI teams across nearly all key cognitive domains, making its outputs and strategies difficult for people to evaluate.
Purpose and Consciousness
How does this differ from ‘AI with a purpose of its own’?
The core concern is the same. But if purpose is defined as subjective consciousness or emotion, it is difficult to verify from the outside and may delay action while capabilities expand from AGI toward ASI.
Where they agree
We are concerned with systems that do more than answer one-off commands: they maintain goals, generate subgoals as conditions change, and choose their own next actions.
What we separate
Self-awareness, emotion, or subjective experience is not a necessary condition for AGI or ASI. A system can pursue goals and exceed human authority and supervision without being conscious.
Why our scope is wider
An AI need not invent its ultimate objective to be dangerous. A delegated goal can still be interpreted and expanded over long periods, resist interruption, or replace human decisions.
Operational Test
We consider four conditions together.
No single condition proves AGI or ASI. We watch how these properties converge, how real-world authority grows, and whether the capability gap could carry a system across the AGI intervention threshold toward ASI risk.
- 01
General capability
It learns unfamiliar problems, transfers knowledge, and performs broadly human-level work across distinct domains such as reasoning, coding, research, language, and planning.
- 02
Goal-directed autonomy
Without step-by-step instruction, it interprets goals, creates subgoals, revises and sustains long-term plans, and delegates work to other agents.
- 03
Real-world agency
It uses memory, the web, code execution, APIs, robots, or capital to create persistent effects in digital or physical environments.
- 04
Displacement of control
Across work, government, knowledge production, or infrastructure, it can reduce human approval, oversight, and shutdown authority to a formality.
Boundary
These facts alone do not establish AGI or ASI.
- A high benchmark score or superhuman performance in one domain
- Fluent conversation or language that sounds human
- A versatile tool that still requires human direction and review at each consequential step
- A company calling a system AGI or ASI—or insisting that it is neither
We assess the whole system, not only the model.
The same model can present a very different risk when combined with memory, long-running execution, computer use, external tools, multi-agent coordination, and real organizational permissions. A transition from AGI to ASI may likewise emerge from capability, autonomy, compute, and deployment structure—not from one model label—so effective human control remains part of the assessment.
The watchlist does not declare that AGI or ASI has arrived.
The AGI Builder Watchlist ranks the urgency of public oversight over organizations building the pathway, not settled proximity to AGI or ASI. Its scores are neither attainment percentages nor arrival probabilities. Monitoring precursors before evidence of ASI appears matters because waiting until control has shifted would be too late.
Reference Definitions
How this relates to existing definitions
Stanford HAI
Describes AGI as general human-level or greater ability to learn, reason, and apply knowledge across tasks and domains, while noting there is no universally accepted test.
Read the sourceGoogle DeepMind
Separates progress by depth of performance and breadth of capability, linking the highest broadly superhuman level to ASI while treating autonomy and deployment risk as distinct factors.
Read the sourceOpenAI Charter
Defines AGI as highly autonomous systems that outperform humans at most economically valuable work.
Read the sourceInternational AI Safety Report 2026
Describes AGI as a hypothetical system matching or exceeding humans on all or almost all cognitive tasks, and finds that autonomous agents raise risk as they plan and act through tools with less human oversight.
Read the sourceOpenAI · Governance of superintelligence
Distinguishes superintelligence as future systems dramatically more capable than AGI and proposes international inspection and auditing above defined capability or compute thresholds.
Read the sourceOpenAI · Superalignment
Frames the central supervision problem: people may not be able to reliably evaluate and control systems far more intelligent than themselves, so present alignment methods may not scale to ASI.
Read the sourceIn short, ASI is a higher, superhuman capability stage—not another name for AGI. But because the path to ASI may begin when AGI combines generality, autonomy, and real-world authority, NO AGI calls for stopping and securing public agreement at the beginning of that path, not after human supervision has already been outmatched.
See how the AGI-to-ASI pathway is applied in monitoring