Guide
What Is ASI?
A plain-English guide to artificial superintelligence: what ASI means, how it differs from AGI, who says they are building it, the forecasts and the safety debate, and why the US government calls ordinary AI 'Super Intelligence'.
Quick answer: ASI, or artificial superintelligence, is a hypothetical AI system that far exceeds the best human minds in virtually every field, from science and strategy to persuasion and invention. It is one step beyond AGI, which means roughly matching humans across most cognitive work. No one claims ASI exists today, but Meta, Microsoft, OpenAI and Safe Superintelligence Inc. all name superintelligence as their goal. The one thing to know before reading any headline: the US government now uses “Super Intelligence” (SI) as its official name for ordinary AI, which is a rebrand of today’s technology, not a claim that ASI has been built.
This page is the explainer: what ASI means, how it could come about, who is working towards it, and the arguments about risk, with every forecast and claim attributed to whoever made it. It takes no position on whether or when ASI will arrive. For the step below ASI, see what is AGI. For all the ways AI is classified, see types of AI. For one-line definitions of everything else, see the AI glossary.
What is ASI?
The standard definition comes from the philosopher Nick Bostrom’s 2014 book Superintelligence: Paths, Dangers, Strategies: superintelligence is “any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest” (Wikipedia). Artificial superintelligence is that intellect built as an AI system.
Two words carry the definition. Super means beyond the best humans, not merely beyond the average person. Intelligence here means general problem-solving ability, so a system that is superhuman at one game does not qualify.
The idea is older than modern AI. In 1965 the statistician I. J. Good wrote, in “Speculations Concerning the First Ultraintelligent Machine”, that “the first ultraintelligent machine is the last invention that man need ever make”, because such a machine could design even better machines (Wikipedia). Bostrom’s book made “superintelligence” the standard term.
AGI vs ASI: what is the difference?
AGI is about breadth at roughly human level. ASI is about level: far beyond the best humans, across almost everything. Narrow AI, the kind in everyday use, is the rung below both.
| Narrow AI | AGI (artificial general intelligence) | ASI (artificial superintelligence) | |
|---|---|---|---|
| What it means | Built for one task or domain | Roughly human-level across most cognitive tasks | Far beyond the best humans across virtually all domains |
| Measured against | The task it was built for | A capable or well-educated adult | The best human experts in every field, combined and exceeded |
| A typical definition | Not applicable | ”Highly autonomous systems that outperform humans at most economically valuable work” (OpenAI Charter) | “Any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest” (Bostrom, 2014) |
| Google DeepMind’s Levels of AGI | Can reach Level 5, Superhuman, within its task | Levels 1 to 4 (Emerging to Virtuoso) of general AI | Level 5, Superhuman, across the board |
| Exists today? | Yes, everywhere: a spam filter, a chess engine, AlphaFold | Openly contested: OpenAI President Greg Brockman greeted GPT-6 Astra with “Welcome to the AGI era” on 3 September 2026, while the ARC Prize Foundation said it lacked “evidence to call this AGI yet” (see what is AGI) | No one claims it does |
Narrow systems can already be superhuman inside their lane. Google DeepMind’s “Levels of AGI” paper (November 2023) lists AlphaFold and AlphaZero as “Superhuman” narrow AI, meaning they outperform every human at their single task. The same table’s “Superhuman” general level, which the paper equates with artificial superintelligence, is marked “not yet achieved”. ASI is superhuman and general at the same time.
The two are linked by one widely discussed idea: that an AGI good enough to do AI research could speed up its own improvement and reach superintelligence quickly. How quickly is one of the biggest disagreements in the field (see the forecasts below).
The boundary between the two is fuzzy in practice. Anthropic’s “powerful AI”, intellectual capability matching or exceeding Nobel Prize winners across most disciplines, described as a “country of geniuses in a datacenter” (quoted by Redwood Research), sits somewhere between the strongest AGI definitions and the weakest ASI ones.
”Super Intelligence” (SI): the US government’s name for AI
Searches for “superintelligence” now catch two different things. One is the technical concept this page is about. The other is a political rebrand of ordinary AI.
| Date | What happened | Source |
|---|---|---|
| 22 Sep 2026 | In his speech to the 81st UN General Assembly, President Donald Trump says the US government will call AI “super intelligence”: “It sounds much better. It is much better, and it’s much more accurate.” | Khaleej Times (AFP); Defense One |
| 29 Sep 2026 | Trump signs the executive order “Inaugurating The Era Of Super Intelligence”. Federal agencies must use “Super Intelligence” and “SI” in place of “AI” in official correspondence, public communications, policy documents and non-statutory documents | White House fact sheet |
| 29 Sep 2026 | Tech executives sign a voluntary “White House Accord on Super Intelligence”; Trump calls it “morally binding” for those who signed | Fox Business |
| Oct 2026 | Elon Musk’s AI company, then called SpaceXAI, is renamed SpaceXSI (SpaceX Super Intelligence) | xAI provider page; Tom’s Hardware |
| 19 Oct 2026 | The SEC announces a roundtable on “Machine-Understandable Data in the Age of Super Intelligence”, set for 9 November 2026 | SEC |
| 5 Nov 2026 | The Department of Justice launches a “Super Intelligence (SI) Task Force” focused on misuse of SI by foreign states and state-sponsored actors, transnational criminal organisations, and people or companies engaged in fraud | US Department of Justice |
| 6 Nov 2026 | Tom’s Hardware reports a Tesla chip roadmap slide that labels the AI4, AI5 and AI6 chips SI4, SI5 and SI6 | Tom’s Hardware |
The White House fact sheet sets the new name against the old one, saying “Super Intelligence” is “not simply imitating or automating human intelligence as is implied by the term ‘Artificial Intelligence’” (fact sheet). The fact sheet gives no definition of SI; the order directs the Assistant to the President for Science and Technology to propose a federal one.
The government label and the technical term mean different things. In official US usage, “SI” now covers the same chatbots, image generators and voice-cloning tools that everyone else calls AI. Announcing charges in an alleged $129 million scheme to impersonate company executives with AI voice clones, US Attorney Jamie McDonald said the accused had “abused groundbreaking super intelligence technology” (Townhall; Associated Press). In the technical sense, superintelligence means a system that beats the best humans at virtually everything, and nobody, including the White House, says such a system exists. Engadget wrote that “most publications have refused to jump on board” with the government’s term.
What forms could superintelligence take?
Bostrom’s book describes three ways a mind could be superintelligent (LessWrong reading-group summary of the book). They are useful for seeing that ASI does not have to look like a single super-genius.
| Form | What it means | Illustration |
|---|---|---|
| Speed superintelligence | A mind like a human’s that simply thinks much faster | ”A human-like reasoner who could think millions of times faster than current humans” (Wikipedia) |
| Collective superintelligence | Many smaller intelligences coordinated well enough to outperform any individual | A large number of separate reasoning systems that communicate and coordinate well enough to act together (Wikipedia) |
| Quality superintelligence | A mind that is qualitatively smarter, not just faster | A system that “can carry out intellectual tasks that humans just can’t in practice” (LessWrong summary) |
How could ASI come about?
There are three main routes in the literature. None has been demonstrated.
- An intelligence explosion. I. J. Good’s 1965 argument: a machine that is better than humans at designing machines could design a better successor, which designs a better one again, a loop known as recursive self-improvement (Wikipedia). This is the root of the idea that ASI could follow AGI quickly.
- Automating AI research. The modern version of Good’s loop. Former OpenAI researcher Leopold Aschenbrenner’s “Situational Awareness” essay (June 2024) argued that by 2027 AI systems would be able to do their own AI research, compressing a decade of algorithmic progress into less than a year (Wikipedia). The AI 2027 scenario, published in April 2025 by former OpenAI researcher Daniel Kokotajlo and four co-authors, describes the same path: AI automates coding, then AI research, and over 2027 goes from doing a research engineer’s job to “eclipsing all humans at all tasks” (AI 2027).
- Continued scaling. The expectation that bigger models, more compute, reinforcement learning and agent tooling will keep raising capability, with no hard ceiling at human level. Whether scaling alone is enough even for AGI is disputed (see what is AGI).
The futurist versions of the same idea go by “the singularity”. The science-fiction author Vernor Vinge wrote in his 1993 essay “The Coming Technological Singularity” that he would be surprised if it came before 2005 or after 2030, and the inventor Ray Kurzweil has predicted human-level AI by 2029 and a singularity by 2045, dates he reaffirmed in his 2024 book The Singularity Is Nearer (Wikipedia).
Who is trying to build superintelligence?
Several of the best-funded AI organisations name superintelligence, rather than AGI, as their goal. Each describes it in its own words.
| Organisation | What it says it is building | Key facts |
|---|---|---|
| Safe Superintelligence Inc. (SSI) | A safe superintelligence: “one goal and one product” (SSI) | Founded on 19 June 2024 by former OpenAI chief scientist Ilya Sutskever with Daniel Gross and Daniel Levy; raised $1 billion at a $5 billion valuation in September 2024 and reached a $30 billion valuation in March 2025; Gross left for Meta Superintelligence Labs in July 2025 and Sutskever became chief executive; on 27 July 2026 announced an Nvidia partnership that included a planned $5 billion investment (Wikipedia) |
| Meta (Meta Superintelligence Labs) | “Personal superintelligence”, in Mark Zuckerberg’s July 2025 letter | Created on 30 June 2025 under chief AI officer Alexandr Wang, after a $14.3 billion investment in Scale AI; released its first model, Muse Spark, on 8 April 2026 and the open-weight Muse Glimmer on 10 August 2026 (Wikipedia) |
| Microsoft (MAI Superintelligence Team) | “Humanist superintelligence” | Announced by Microsoft AI chief Mustafa Suleyman on 6 November 2025; the application areas he named were an AI companion for everyone, medical superintelligence and plentiful clean energy (Microsoft AI) |
| OpenAI | Superintelligence, beyond AGI | Sam Altman wrote in January 2025: “We are beginning to turn our aim beyond that, to superintelligence in the true sense of the word” (Reflections; Maginative) |
Two other leading labs frame their goal differently. Google DeepMind describes its mission in terms of AGI, and Anthropic uses “powerful AI”. Neither label is a claim about superintelligence.
OpenAI’s leaders wrote about superintelligence well before this. In May 2023 Altman, Greg Brockman and Ilya Sutskever published “Governance of superintelligence”, which said it was “conceivable that within the next ten years, AI systems will exceed expert skill level in most domains” and that “we are likely to eventually need something like an International Atomic Energy Agency (IAEA) for superintelligence efforts” (Business Standard).
When could ASI arrive? The forecasts
There is no consensus on whether ASI is possible, let alone when. The forecasts below are reported, not endorsed.
| Forecast | Who | Estimate | Source |
|---|---|---|---|
| Is the intelligence-explosion argument broadly correct? | 2017 email survey of authors who published at the 2015 NeurIPS and ICML conferences | 12% “quite likely”, 17% “likely”, 21% “about even”, 24% “unlikely”, 26% “quite unlikely” | Wikipedia |
| AI doing its own AI research | Leopold Aschenbrenner, “Situational Awareness” (2024) | By 2027, followed by rapid progress towards superintelligence | Wikipedia |
| AI surpassing all humans at all tasks | AI 2027 scenario (2025) | During 2027 in the scenario, though the authors say their median forecasts were “somewhat longer” | AI 2027 |
| Singularity | Ray Kurzweil | 2045 (human-level AI by 2029) | Wikipedia |
| Singularity | Vernor Vinge (1993) | Between 2005 and 2030 | Wikipedia |
The expert survey on the intelligence explosion is from 2017, before today’s large language models existed, and it showed a field split almost evenly. Most of the other forecasts depend on AGI arriving first.
For calendar forecasts of AGI itself, the step that would come first, see the AGI forecasts table.
Is superintelligence dangerous?
This is where the debate is loudest, and credible people sit at both ends of it.
The case for serious concern
The core worry is control. A system much smarter than its builders could pursue goals they did not intend, and might be impossible to correct once deployed. The computer scientist Roman Yampolskiy’s 2024 book is titled AI: Unexplainable, Unpredictable, Uncontrollable, and Eliezer Yudkowsky and Nate Soares’s 2025 book If Anyone Builds It, Everyone Dies is subtitled Why Superhuman AI Would Kill Us All (Wikipedia).
The labs building towards it say the problem is open. In an October 2026 interview with Fortune, Sam Altman said “We have not solved alignment” and “I believe no lab has solved alignment” (Fortune).
The most visible public campaign is the Statement on Superintelligence, hosted by the Future of Life Institute and launched on 22 October 2025 (SiliconANGLE). It reads: “We call for a prohibition on the development of superintelligence, not lifted before there is broad scientific consensus that it will be done safely and controllably, and strong public buy-in.” Signatories include AI pioneers Geoffrey Hinton, Yoshua Bengio and Stuart Russell, Apple co-founder Steve Wozniak, Richard Branson, Steve Bannon, Susan Rice, and Prince Harry and Meghan (FLI). The statement had more than 78,000 signatures in November 2026. A US poll FLI released with it found 64% of respondents agreed that superintelligence should not be developed until there is scientific agreement that it is safe and controllable (FLI; SiliconANGLE).
The idea has also reached legislation. In September 2026 Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act, which would pause advanced AI development until a cabinet-level Department of Artificial Intelligence is in place to oversee it (Wikipedia; The Debrief, 25 September 2026).
The case for scepticism
The opposite camp argues that superintelligence is either far off, poorly defined or not a useful way to think about AI at all.
- “Superintelligence” is the wrong frame. The computer scientists Arvind Narayanan and Sayash Kapoor, in their April 2025 essay “AI as Normal Technology”, describe “superintelligent” AI as “incoherent as usually conceptualized” and argue that AI is a normal technology, like electricity and the internet, whose transformative effects will be slow, “on the timescale of decades” (Knight First Amendment Institute).
- Progress follows S-curves. Stuart Russell and Peter Norvig note that “improvement in a particular area tends to follow an S curve”, flattening out rather than accelerating forever, and the economist Robert J. Gordon argues the economic data show no trace of a coming singularity (Wikipedia).
- Imagining it is not evidence for it. Steven Pinker wrote in 2008: “There is not the slightest reason to believe in a coming singularity. The fact that you can visualize a future in your imagination is not evidence that it is likely or even possible” (Wikipedia).
- Today’s methods are not enough. Yann LeCun, who left Meta in November 2025 and founded AMI Labs, argues that large language models lack the world models needed even for human-level AI (see what is AGI).
The two camps agree on more than it looks. Almost nobody argues that today’s systems are superintelligent. The disagreement is about whether the path from here leads there, how fast, and how much caution that justifies now.
How to read a superintelligence claim
When a company, politician or headline uses the word, four questions sort it out.
- Is it the technical term or the government label? Since September 2026, “Super Intelligence” or “SI” in a US government document means ordinary AI.
- Which definition? Bostrom’s “virtually all domains” is a very high bar. A system that is superhuman at coding or chess is narrow, however impressive.
- Is it a goal or a claim? “We are building superintelligence” (Meta, Microsoft, SSI) is a mission statement. No company says it has one.
- Who benefits if it is believed? Superintelligence language features in fundraising, hiring and policy fights. That does not make a claim wrong, but it is context worth having.
For the practical question of which AI model is strongest today, the superintelligence debate changes nothing. That is answered by task-specific evidence: see best AI models and best AI agents.
Related reading
For the step below ASI, the definitions, benchmarks and the GPT-6 Astra debate, see what is AGI. For narrow AI, AGI, ASI and the other ways AI is classified, types of AI. For how today’s frontier systems work, what is an LLM and what is agentic AI. For one-line definitions of superintelligence, alignment and 100-plus other terms, the AI glossary. For the companies named on this page, our provider pages for OpenAI, Meta, Microsoft, Google, Anthropic and xAI, and the GPT-6 Astra model page.
Frequently asked questions
What is ASI in simple terms?
ASI, or artificial superintelligence, is a hypothetical AI that would be far smarter than the best humans at almost everything, not just at one task. Today’s AI can beat humans at specific things, such as chess or protein-structure prediction, but no system is superhuman across the board. No one claims ASI exists.
What does ASI stand for?
ASI stands for artificial superintelligence. It is the third rung in the common three-level scheme of AI capability: narrow AI (built for specific tasks), artificial general intelligence or AGI (roughly human-level across most tasks), and artificial superintelligence or ASI (far beyond human level across virtually all tasks).
What is superintelligence?
Superintelligence is any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest, in the definition from Nick Bostrom’s 2014 book Superintelligence. When the intellect is an AI system, it is called artificial superintelligence (ASI). The idea dates back at least to I. J. Good’s 1965 paper on an “ultraintelligent machine”.
What is the difference between AGI and ASI?
AGI (artificial general intelligence) means AI that matches human ability across most cognitive tasks. ASI (artificial superintelligence) means AI that greatly exceeds the best humans across virtually all of them. Whether AGI has arrived is openly contested, especially since OpenAI’s GPT-6 Astra launched in September 2026; no one claims ASI has. See what is AGI.
Does artificial superintelligence exist yet?
No. No company, researcher or government claims that artificial superintelligence exists. Some narrow AI systems, such as AlphaFold and AlphaZero, are superhuman at a single task, but superintelligence requires superhuman ability across virtually every domain. Meta, Microsoft, OpenAI and Safe Superintelligence Inc. describe it as a goal.
Why does the US government call AI “super intelligence”?
President Donald Trump announced at the UN General Assembly on 22 September 2026 that the US government would call AI “super intelligence”, and signed the executive order “Inaugurating The Era Of Super Intelligence” on 29 September 2026. Federal agencies now use “Super Intelligence” and “SI” in place of “AI” in official communications. The White House says the old name implies technology that merely imitates or automates human intelligence; the change is a rebrand of today’s AI, not a claim that technical superintelligence exists.
What does SI stand for in AI?
SI stands for Super Intelligence, the official US government name for artificial intelligence since an executive order signed on 29 September 2026. Agencies such as the Department of Justice, which launched a Super Intelligence Task Force on 5 November 2026, and the SEC now use it. Elon Musk’s AI company was renamed SpaceXSI in October 2026, and a Tesla chip roadmap reported in November 2026 labels its chips SI4, SI5 and SI6. In this usage SI means ordinary AI, not ASI.
When will ASI arrive?
No one knows, and some researchers doubt it is possible. Leopold Aschenbrenner argued in 2024 that AI would be doing its own AI research by 2027, with superintelligence following quickly, and Ray Kurzweil has predicted a technological singularity by 2045. In a 2017 survey of machine-learning researchers, half rated the intelligence-explosion argument “unlikely” or “quite unlikely”. Sceptics such as Arvind Narayanan and Sayash Kapoor describe superintelligence as incoherent as usually conceived.
Is superintelligence dangerous?
Many leading researchers think it could be. The Statement on Superintelligence, launched by the Future of Life Institute in October 2025 and signed by Geoffrey Hinton, Yoshua Bengio and tens of thousands of others, calls for a prohibition on developing superintelligence until there is broad scientific consensus it can be done safely and controllably, and strong public buy-in. OpenAI CEO Sam Altman said in October 2026 that no lab has solved alignment. Sceptics argue that superintelligence is poorly defined or far off, and that AI’s effects will arrive gradually.
Who is building superintelligence?
Meta (Meta Superintelligence Labs, created June 2025), Microsoft (the MAI Superintelligence Team, announced November 2025), OpenAI (which said in January 2025 it is turning its aim to superintelligence) and Safe Superintelligence Inc. (founded by Ilya Sutskever in June 2024) all name superintelligence as their goal. None says it has built one.
What is Safe Superintelligence Inc.?
Safe Superintelligence Inc. (SSI) is an AI company founded on 19 June 2024 by former OpenAI chief scientist Ilya Sutskever, Daniel Gross and Daniel Levy, with offices in Palo Alto and Tel Aviv. It describes its aim as “one goal and one product: a safe superintelligence”. It raised $1 billion in September 2024, reached a $30 billion valuation in March 2025, and announced an Nvidia partnership including a planned $5 billion investment in July 2026.
What is the technological singularity?
The technological singularity is a hypothetical point at which technological growth, usually driven by superintelligent AI improving itself, becomes so fast that it is beyond human control or prediction. Vernor Vinge wrote in his 1993 essay “The Coming Technological Singularity” that he would be surprised if it came before 2005 or after 2030; Ray Kurzweil predicts it by 2045.
Can superintelligence be controlled?
There is no agreed answer. Keeping a far smarter system aligned with human goals is an unsolved research problem: Sam Altman said in October 2026 that “no lab has solved alignment”, and the computer scientist Roman Yampolskiy’s 2024 book calls AI “uncontrollable” in its title. Others argue superintelligence is far enough off, or poorly enough defined, that the question is premature.