Reality check
ASI is hypothetical, and no webpage can certify it
An artificial superintelligence would not merely chat fluently or beat a champion. It would broadly and substantially exceed the best human capabilities. This audit helps structure evidence for a candidate system you have in mind. Your answers are inputs, not verified facts; the result identifies evidence gaps rather than declaring ASI.
Evidence reviewed August 2026
Interactive claim stress test
ASI Evidence & Assurance Audit
Choose answers for one candidate system or claim. If evidence is private, anecdotal or unknown, select “Not verified.”
Progress0/10
How to answer: “Verified” requires documented, repeatable evidence; “Partial” means promising but limited; “Not verified” includes unknown, merely claimed or untested. Do not award points for confidence, branding or forecasts.
0/20evidence points
Candidate evidence profile
What this result means
It is not proof of ASI. The audit did not inspect the candidate, validate your answers or establish consciousness, benevolence or deployment safety.
Capability and assurance are not interchangeable. Strong safeguards cannot turn a weak system into ASI. Strong capability without assurance is a warning, not a deployment recommendation.
Define the extraordinary claim
What is artificial superintelligence?
Broad superiority, not a single superpower
ASI usually means artificial intelligence that greatly exceeds the best humans across essentially all or most relevant cognitive domains: scientific discovery, engineering, strategic planning, communication, creativity and learning itself.
The word greatly matters. Matching a typical person broadly is closer to some AGI definitions. Beating every person at one game is superhuman narrow AI. ASI combines breadth with superhuman depth.
Operational claim: name the domains, human comparison groups, margin of superiority, duration, system resources and independent verification standard.
Narrow AIsuperhuman specialty
breadth + depth
AGIbroad human-range
→
ASIbroadly beyond humanity
Do not skip levels
Four very different capability claims
Exceeds all humans in a bounded task such as chess or a specialized prediction problem.
Works across many cognitive tasks but remains uneven, brittle or below skilled humans.
Broadly meets or exceeds skilled-human performance under a stated definition.
Broadly and substantially outperforms the best human individuals or teams.
| Observation | What it supports | What it does not establish |
|---|---|---|
| Wins at a game | Superhuman task-specific performance | Broad generality |
| Passes a Turing-style test | Human-like conversation in that protocol | ASI or consciousness |
| Scores above most humans on exams | Broad test performance | Novel research, autonomy or best-human superiority |
| Completes agent tasks | Some planning and tool-use ability | Reliable long-horizon agency everywhere |
| Accelerates AI research | Potential feedback into capability development | An inevitable intelligence explosion |
The idea before the evidence
History of artificial superintelligence
1950
Chapter 01
Machine intelligence becomes operational
Turing’s imitation game focuses debate on behavior, decades before “superintelligence” becomes a common technical-policy term.
1965
Chapter 02
I. J. Good’s ultraintelligent machine
Good defines an ultraintelligent machine as surpassing all human intellectual activity and argues that designing better machines could trigger an intelligence explosion.
1993
Chapter 03
Vinge’s technological singularity
Vernor Vinge describes the creation of greater-than-human intelligence as a discontinuity after which familiar models of progress may fail.
1997–2016
Chapter 04
Narrow systems surpass champions
Deep Blue, Watson and AlphaGo demonstrate superhuman domain performance—important milestones that are not themselves ASI.
2014
Chapter 05
Superintelligence enters wider debate
Nick Bostrom’s book organizes pathways, control problems and strategic questions around intelligence far beyond human levels.
2023
Chapter 06
Frontier governance proposals expand
Labs and researchers publish proposals for superintelligence governance, scalable oversight and capability-triggered safety measures.
2024–26
Chapter 07
Responsible scaling and frontier evaluations
Organizations increasingly tie safeguards to dangerous-capability thresholds, while policies, definitions and commitments continue to evolve.
Future
Chapter 08
ASI remains hypothetical, not certified
No universally recognized body has announced or verified an artificial superintelligence. Claims must be evaluated against explicit breadth, depth and assurance standards.
Multiple routes, uncertain speeds
How ASI might emerge in theory
Intelligence explosion
A system improves AI research, enabling more capable successors that improve the process again. Bottlenecks in hardware, data, experiments and coordination could limit the loop.
Collective machine intelligence
Large populations of agents, specialist models and tools could together exceed human institutions without one monolithic “mind.”
Broad capability growth
Models may steadily improve across domains until the combined system crosses a superhuman threshold—without one dramatic overnight event.
Risk is not one story
Why ASI could be transformative—and dangerous
Potential benefits
- Accelerated medicine and materials science
- Better climate, energy and infrastructure design
- Highly capable education and accessibility tools
- Automation of dangerous or exhausting work
- New scientific theories and engineering solutions
Potential harms
- Catastrophic misuse in cyber or biological domains
- Loss of meaningful human control
- Concentration of economic and political power
- Manipulation, surveillance and strategic instability
- Large-scale accidents from misspecified objectives
Misuse
A capable system may amplify harmful intentions without itself having hostile goals.
Misalignment
A system can competently optimize an objective that diverges from what people intended.
Institutional failure
Competition, secrecy or weak oversight can turn manageable technical risk into systemic risk.
Unequal control
Even beneficial capability can undermine legitimacy if a few actors determine access and goals.
Defense in depth
No single safety technique is enough
- 01
Set capability thresholds in advance
Define which results trigger stronger containment, security and oversight before they appear.
- 02
Use independent evaluators
Separate those rewarded for shipping a system from those empowered to challenge the evidence.
- 03
Test for dangerous affordances
Evaluate cyber, biological, persuasion, autonomy and AI-research capabilities with appropriate controls.
- 04
Limit permissions by default
Network, code, money, replication and physical systems should require explicit, monitored authorization.
- 05
Plan incidents and shutdowns
Logging, rollback, containment, notification and clear decision authority should exist before deployment.
- 06
Build public legitimacy
Systems with cross-border consequences require more than private terms of service or voluntary promises.
Three futures, not one prophecy
Useful scenarios for thinking clearly
Cooperative acceleration
Advanced systems remain controlled, distribute gains widely and accelerate science under legitimate oversight.
Uneven transformation
Capabilities advance quickly but institutions adapt slowly, producing major productivity gains alongside disruption and concentrated power.
Loss of control
Misuse, misalignment or strategic competition outruns safeguards, causing severe or irreversible harm.
Scenario thinking is not probability assignment. Its purpose is to reveal assumptions, early warning indicators and decisions that remain useful across several possible futures.
Clear answers
ASI frequently asked questions
What does ASI stand for?
Artificial superintelligence: a hypothetical artificial system that broadly and substantially exceeds the best human cognitive performance, not merely average human skill or one specialist record.
Is ASI the same as AGI?
No. AGI generally means broad, adaptable capability around human levels or above. ASI is a stronger concept: broad performance far beyond human capability. The boundary depends on the operational framework.
Is there an official ASI Test?
No. There is no universal authority, benchmark or legal certificate. This page offers a transparent evidence-and-assurance rubric, not an official certification.
Does my score prove a system is ASI?
No. Your score summarizes the evidence status you selected. The page cannot verify your inputs or inspect a live system. It helps reveal missing evidence and governance gaps.
Can a chess engine be ASI?
No under broad definitions. It may be superhuman at chess while remaining narrow. ASI requires superhuman breadth as well as depth.
Would ASI have to be conscious?
Not under capability-based definitions. Consciousness, moral status and subjective experience are separate unresolved questions.
What is an intelligence explosion?
It is the hypothesis that a sufficiently capable machine could improve AI research, enabling better successors and creating a rapid positive-feedback cycle in capability. Its speed and feasibility are debated.
What is recursive self-improvement?
A system contributes to improving the algorithms, data, hardware use or research processes that produce more capable versions, which may in turn improve the process further.
Could ASI be slow rather than sudden?
Yes. Superhuman breadth could emerge through gradual accumulation, coordinated systems or economic scaling rather than a single abrupt event. “ASI” does not logically require one specific takeoff speed.
Does high capability imply harmful intent?
No. Capability does not determine goals. Risk can arise from misuse, error, poorly specified objectives, power-seeking instrumental behavior, concentration of control or institutional failure.
Does alignment solve every ASI risk?
No. Even a well-intentioned system can be misused, stolen, deployed recklessly or embedded in unjust institutions. Technical alignment, security and governance are complementary.
Why include assurance gates in an ASI test?
A capability classification and a deployment decision are different. Assurance cannot make a weak system superintelligent, but extraordinary capability without adequate control evidence would be especially dangerous.
Has ASI already been achieved?
No universally accepted, independently verified ASI has been established. Narrow superhuman systems and increasingly general models should not be relabeled ASI without broad evidence.
What would be the clearest evidence?
Repeated, independent demonstrations of broad superiority over top human teams on genuinely novel, consequential work—combined with transparent costs, system boundaries and failure analysis.
Trace the argument
Primary and research sources
Good (1965): First Ultraintelligent Machine
Foundational discussion of an ultraintelligent machine and intelligence explosion.
Open sourceVinge (1993): Technological Singularity
NASA-hosted record of the essay on greater-than-human intelligence and discontinuity.
Open sourceGoogle DeepMind: Levels of AGI
Places ASI at broad superhuman performance and separates breadth from depth.
Open sourceOpenAI: Governance of Superintelligence
A lab-authored proposal discussing systems dramatically more capable than AGI and public oversight.
Open sourceManaging Extreme AI Risks
Research paper on misuse, alignment, autonomous systems and governance mechanisms.
Open sourceNIST AI Risk Management Framework
General framework for governing, mapping, measuring and managing AI risks.
Open sourceInternational AI Safety Report
Internationally authored assessment of general-purpose AI capabilities, risks and mitigations.
Open sourceAnthropic Responsible Scaling Policy
An evolving example of capability thresholds connected to stronger safeguards.
Open source