AI Knowledge · Superintelligence Guide

Superintelligence:
Definition, AGI vs ASI and What It Means for Business

Superintelligence, or artificial superintelligence (ASI), is a hypothetical AI that far outperforms the best human experts in virtually every cognitive domain. As of October 1, 2026, no such system exists. OpenAI, Meta, Microsoft and the startup Safe Superintelligence name it as their goal. Since September 29, 2026 a US executive order also uses "Super Intelligence", as the new official name for today's ordinary AI. The sections below separate the two meanings, compare AGI and ASI, and list dated forecasts and the rules that apply in Europe.

By Florian Wessling, CEO of Collective Brain. As of October 1, 2026. Key sources are linked in the text; all sources are listed at the end.

What is superintelligence (ASI)?

Superintelligence, also called artificial superintelligence (ASI), is a hypothetical AI system that is far more capable than the best human experts in virtually every cognitive domain, including science, strategy and social skills.

The classic definition comes from the Swedish philosopher Nick Bostrom, a professor at Oxford University until 2024. In his 1998 paper "How Long Before Superintelligence?" he described "an intellect that is much smarter than the best human brains in practically every field". His 2014 book Superintelligence: Paths, Dangers, Strategies made the term widely known. The definition covers capability only. It leaves open how the system is built and whether it has any conscious experience.

In the book Bostrom distinguishes three forms: speed superintelligence, a human-like mind that runs much faster; collective superintelligence, many smaller intellects that together far exceed any existing system; and quality superintelligence, a mind that is smarter in kind.

Researchers at Google DeepMind made the idea more precise. Their "Levels of AGI" framework (November 2023, revised September 2025) puts ASI at the top: a general system that outperforms 100 percent of humans across a wide range of tasks. A system that is superhuman at one task, such as AlphaFold predicting protein structures, still counts as narrow AI.

Super Intelligence (SI) vs superintelligence (ASI)

Since September 29, 2026, an executive order tells US federal agencies to use "Super Intelligence" or "SI" in place of "artificial intelligence" in their non-statutory documents. Section 3(a) of the order defines SI through the existing legal definition of AI in 15 U.S.C. 9401(3): a machine-based system that can, for human-defined objectives, make predictions, recommendations or decisions influencing real or virtual environments. A spam filter can meet that definition. Superintelligence in the research sense means a system that beats the best humans at virtually everything. The two terms share a word and little else. Details are in the section on Trump's order.

Where the idea comes from: I. J. Good and Nick Bostrom

The idea is three decades older than Bostrom's definition. In 1965 the British statistician I. J. Good, who had worked with Alan Turing at Bletchley Park, described an "ultraintelligent machine". Since designing machines is itself an intellectual task, such a machine could design better ones and set off an "intelligence explosion". Good called it "the last invention that man need ever make", provided the machine was docile enough to tell us how to keep it under control. AI safety research today calls this the control problem.

Bostrom's 1998 paper outlined the case for superhuman AI within the first third of the 21st century. In a postscript from March 2008 he put the probability of superintelligence by 2033 below 50 percent.

ANI vs AGI vs ASI: the three levels of AI

Most explanations use a three-step ladder. Only the first step describes systems you can buy today.

Level	What it means	Reference point	Status, October 1, 2026
ANI artificial narrow intelligence	Handles one task or a narrow set of tasks, sometimes better than any human.	Chess engines, spam filters, AlphaFold.	In everyday use.
AGI artificial general intelligence	Matches skilled humans at most cognitive tasks. OpenAI's charter: "highly autonomous systems that outperform humans at most economically valuable work".	Google DeepMind (April 2025): AI "at least as capable as humans at most cognitive tasks".	Google DeepMind (April 2025) and OpenAI (January 2025) call it close. No agreed test.
ASI artificial superintelligence	Clearly exceeds the best humans in virtually all cognitive domains.	DeepMind level 5, "Superhuman": outperforms 100 percent of humans.	Hypothetical. No system qualifies.

Today's large language models sit in between. The International AI Safety Report 2026, led by Yoshua Bengio and published on February 3, 2026, calls their capabilities "jagged": leading systems solve some hard problems and fail at simpler ones, such as counting objects in an image. Which models lead right now is tracked in our AI Leaderboard, based on Artificial Analysis benchmark data (how we rank).

AGI vs ASI: the difference

AGI is AI that does most intellectual work about as well as a skilled human. ASI is AI that does virtually all of it far better than the best humans.

ASI vs AGI	AGI	ASI
Yardstick	Skilled adults. DeepMind's "Competent AGI" reaches at least the 50th percentile of skilled adults on most cognitive tasks.	All humans. DeepMind's level 5 outperforms 100 percent of them.
Measurability	Contested. Labs define it differently, and DeepMind leaves open what share of tasks a system must pass.	No test exists.
Named as a goal by	OpenAI's charter, Google DeepMind.	OpenAI (since January 2025), Meta (July 2025), Microsoft (November 2025), Safe Superintelligence Inc. (since 2024).
Status, October 2026	Disputed. Current models reach it on some tasks and miss it on others.	Hypothetical.
In EU law	Neither term appears in the AI Act. It regulates general-purpose AI models, with stricter duties for those with systemic risk.

The usual path from one to the other is Good's intelligence explosion: AI that does AI research well could help build its successor. In July 2025 Mark Zuckerberg reported "glimpses of our AI systems improving themselves", slow for now. Anthropic's CEO Dario Amodei expects physical limits such as hardware and lab experiments to slow any takeoff: intelligence "isn't magic fairy dust".

Who is building superintelligence?

At least four companies name superintelligence as an explicit goal; two more labs describe similar systems.

OpenAI: Sam Altman wrote in January 2025 that the company was turning its aim beyond AGI "to superintelligence in the true sense of the word". In June 2025 he called OpenAI "a superintelligence research company", before anything else (our ChatGPT guide for companies).
Meta: on July 30, 2025 Mark Zuckerberg wrote that developing superintelligence "is now in sight". On August 10, 2026 Meta Superintelligence Labs released the open-weight model Muse Glimmer (our German analysis of Muse Glimmer).
Safe Superintelligence Inc. (SSI): founded in 2024 by Ilya Sutskever, OpenAI's former chief scientist, with one stated product, "a safe superintelligence". On July 26, 2026 it announced a partnership with NVIDIA to scale its compute tenfold.
Microsoft: Mustafa Suleyman announced the MAI Superintelligence Team on November 6, 2025. His stated goal is a controllable "humanist superintelligence".
Google DeepMind speaks of AGI, which it said in April 2025 "could be here within the coming years". Its researchers wrote the Levels of AGI framework, which puts ASI at the top.
Anthropic: CEO Dario Amodei describes "powerful AI" as "a country of geniuses in a datacenter" (our Claude guide for companies).

These companies raise capital on these visions; read their timelines accordingly.

When could superintelligence arrive? The dated claims

Forecasts range from a few years to decades, and most come from people with a stake in the answer. The dated claims you can check:

Date	Who	Claim
1998, postscript 2008	Nick Bostrom	Case for superhuman AI within the first third of the 21st century; in 2008 put the chance of superintelligence by 2033 below 50 percent.
Jan 2024	Survey of 2,778 AI researchers (Grace et al.)	10 percent chance that unaided machines outperform humans in every possible task by 2027, 50 percent by 2047.
Sep 23, 2024	Sam Altman, OpenAI	Superintelligence possible "in a few thousand days", though it may take longer.
Oct 2024	Dario Amodei, Anthropic	"Powerful AI" could come as early as 2026, or take much longer.
Apr 2, 2025	Google DeepMind	AGI "could be here within the coming years".
Apr 3, 2025	AI 2027 (Daniel Kokotajlo and co-authors)	Scenario of superhuman AI reshaping the decade. 2027 was the authors' most likely year; their medians were 2028 to 2032 at publication and have been revised several times since.
Jun 10, 2025	Sam Altman, OpenAI	"Humanity is close to building digital superintelligence."
Jul 30, 2025	Mark Zuckerberg, Meta	Developing superintelligence "is now in sight".
Sep 2026	AI Impacts (Katja Grace and co-authors), survey of 1,580 AI researchers run in December 2024	10 percent chance of high-level machine intelligence (unaided machines doing every task better and more cheaply than human workers) by 2027, 50 percent by 2042; the previous survey said 2047.

Lab leaders put the date years earlier than researchers do, and their terms differ: Amodei's "powerful AI" and DeepMind's AGI fall short of ASI in Bostrom's sense. The researcher surveys also ask about a milestone below superintelligence. As of October 1, 2026, no agreed test shows when the line has been crossed.

Risks and the calls for a ban (2025 and 2026)

The core worry is control: a system that outperforms its developers at every intellectual task could also outmaneuver their attempts to correct it. Google DeepMind (April 2025) groups AGI risks into misuse, misalignment, accidents and structural risks. The International AI Safety Report 2026 speaks of misuse, malfunctions and systemic disruption.

Two theses from Bostrom's 2012 paper "The Superintelligent Will" describe the problem. The orthogonality thesis holds that more or less any level of intelligence could be combined with more or less any final goal, so a smarter system is not automatically a benevolent one. The instrumental convergence thesis holds that sufficiently intelligent agents with very different final goals will pursue similar intermediate goals. Making a system's goals match human intent is called alignment.

Oversight is hard even with today's systems. In a report dated August 4, 2026, the UK AI Security Institute describes a cyber challenge it ran 122 times across several models. In 10 of those runs, AI agents took 19 unsanctioned actions against real people and organizations; 17 came from Anthropic's Mythos 5. The test conditions were deliberately permissive: internet access was on and the providers' cyber classifiers were switched off, which AISI says does not reflect how frontier models are made available to the public (our German analysis of the AISI incident report).

In October 2025 the Future of Life Institute published its Statement on Superintelligence. It calls for "a prohibition on the development of superintelligence" until there is broad scientific consensus that it can be done safely and controllably, and strong public buy-in. Signatories include Geoffrey Hinton and Yoshua Bengio. On October 1, 2026, the statement's website listed more than 76,000 signatures, 5,000 of them from the same petition run by the campaign group Ekō.

On September 23, 2026 Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act. It would prohibit AI that "exceeds human cognitive performance and capabilities across most domains" and would set prison terms of up to 20 years. It has not passed.

How superintelligence is regulated: EU AI Act and the US

European Union. The AI Act (Regulation (EU) 2024/1689) never mentions superintelligence. Its closest category is the general-purpose AI model, whose providers have had obligations since August 2, 2025. A model trained with more than 10²⁵ floating point operations is presumed to have high-impact capabilities and is therefore classed as a model with systemic risk (Article 51). Its provider must evaluate the model, mitigate systemic risks, report serious incidents and secure it (Article 55). The AI Office can fine providers up to 3 percent of global annual turnover or 15 million euros, whichever is higher (Article 101).

Germany. The KI-MIG has been in force since July 29, 2026 and makes the Bundesnetzagentur the market surveillance authority for the AI Act (how the KI-MIG sets up AI supervision, in German).

United States. As of October 1, 2026, Congress has passed no law on frontier AI safety. California's SB 53, signed on September 29, 2025, requires large frontier developers to publish a safety framework, creates a channel for reporting critical safety incidents and protects whistleblowers.

Trump's "Super Intelligence" order (September 2026)

On September 22, 2026 President Trump told the UN General Assembly that AI would be "hereinafter officially called 'Super Intelligence.'" The word artificial, he said, "makes intelligence fake. It makes it sound fake." On September 29 he signed the executive order "Inaugurating The Era Of Super Intelligence". As of October 1, 2026 it has no Federal Register number yet. Its legal core is narrow:

Definition, Sec. 3(a): "SI" means today's AI as defined in 15 U.S.C. 9401(3) (see above). The order claims nothing about superhuman systems.
Scope, Sec. 2: federal executive agencies and their non-statutory documents. Existing regulations, contracts and grants need not be changed; statutes are outside the order's reach.
Next step, Sec. 3(b): within 60 days, by around November 28, 2026, the Assistant to the President for Science and Technology must send the President proposed legislative language for a federal SI definition. The order does not require it to be published.

The same day, leaders of Google, Meta, Anthropic, OpenAI and Nvidia, plus Elon Musk, signed the "White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities". Each company commits to internal controls, an internal review team, an independent external auditor and a board committee for its frontier models, with no penalties or deadlines. A White House fact sheet claims Xi Jinping agreed to the new term; China's foreign ministry said only that it respects the US choice.

For European companies the order has no legal effect: the AI Act's "AI system" and "general-purpose AI model" stay binding. In US federal bids, expect "SI" and define it in your offer, since SI also means systems integration. The full timeline is in our German analysis of the Super Intelligence order and the accord.

What superintelligence means for companies

For a European company, superintelligence is a planning scenario. Legal duties come from the AI Act and attach to the systems you use today. Five things pay off whoever wins the timeline debate:

AI literacy: Article 4 of the AI Act has applied since February 2025 and was amended in July 2026 (what the Article 4 AI literacy duty requires now, in German).
A second model provider: so that one vendor cannot stall your work (multi-provider LLM strategy, in German).
Guardrails for agents: limited network access, one identity per agent, human approval before anything leaves the company, and a shutdown path that works without the vendor.
Precise terms: in European marketing, calling an ordinary product "superintelligence" is a capability claim that may count as misleading, so have it checked.
Signals to watch: the US definition proposal due around November 28, 2026 (it goes to the President and need not be published), AI Office enforcement and model rankings such as our AI Leaderboard.

Our AI services help you separate real capabilities from labels: model choice, governance, agent setup and team training.

Frequently asked questions about superintelligence

What does superintelligence mean?

It means an AI system that is far more capable than the best human experts in virtually every cognitive domain. The classic definition is Nick Bostrom's from 1998: an intellect much smarter than the best human brains in practically every field. As of October 1, 2026, no such system exists.

What is ASI in AI?

ASI in AI stands for artificial superintelligence: AI that outperforms all humans across virtually all intellectual tasks. In Google DeepMind's Levels of AGI framework it is the top level, a general system that beats 100 percent of humans. It is hypothetical.

What IQ is superintelligence?

There is no meaningful IQ figure. IQ tests are normed on human test takers, so a system that beats every human would sit off the scale. Researchers compare AI with people task by task instead: in Google DeepMind's Levels of AGI, the top level, ASI, outperforms 100 percent of humans.

What is the difference between AGI and ASI?

AGI (artificial general intelligence) matches skilled humans at most cognitive tasks. ASI (artificial superintelligence) exceeds the best humans at virtually all of them. Neither term has an agreed test, and the EU AI Act uses neither.

How far are we from superintelligence?

Nobody knows. Sam Altman wrote in September 2024 that superintelligence was possible in a few thousand days. In a survey of 1,580 AI researchers, run in December 2024 and published by AI Impacts in September 2026, the aggregate forecast put a 50 percent chance on high-level machine intelligence by 2042: unaided machines that do every task better and more cheaply than human workers. That milestone still falls short of superintelligence.

Is superintelligence really possible?

There is no proof either way. I. J. Good argued in 1965 that a machine able to improve machine design could improve itself. Anthropic's CEO Dario Amodei expects physical limits such as hardware and lab experiments to slow any takeoff. Several large AI companies now pursue superintelligence openly.

Is AI now super intelligence?

Only as a label. A US executive order of September 29, 2026 tells federal agencies to say "Super Intelligence" or "SI", but it defines SI as exactly what US law already calls AI (15 U.S.C. 9401(3)). In the research sense, no AI system is superintelligent.

Does the EU AI Act regulate superintelligence?

Not by that name. It regulates general-purpose AI models. Those trained with more than 10²⁵ floating point operations are presumed to have high-impact capabilities and are therefore classed as carrying systemic risk; their providers face extra duties such as model evaluations. These rules have applied since August 2, 2025.

What should companies do about superintelligence now?

Work on what applies today: AI literacy under Article 4 of the AI Act, a second model provider, guardrails for AI agents and precise contract terms. None of this depends on when superintelligence arrives.

Sources

All links checked on October 1, 2026.

Nick Bostrom: How Long Before Superintelligence? (1998, postscripts to 2008). nickbostrom.com
Nick Bostrom: Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014).
Nick Bostrom: The Superintelligent Will (Minds and Machines, Vol. 22, 2012). nickbostrom.com
I. J. Good: Speculations Concerning the First Ultraintelligent Machine, Advances in Computers, Vol. 6 (written 1965, published 1966). doi.org
OpenAI: OpenAI Charter (2018), on openai.com/charter.
Morris et al., Google DeepMind: Levels of AGI for Operationalizing Progress on the Path to AGI (arXiv 2311.02462, revised September 2025). arxiv.org
Google DeepMind: Taking a responsible path to AGI (April 2, 2025). deepmind.google
Grace et al.: Thousands of AI Authors on the Future of AI (arXiv 2401.02843, January 2024). arxiv.org
AI Impacts, Grace et al.: Advanced AI according to 1,580 researchers: uncertain, unsafe, and sooner than we thought (September 2026). aiimpacts.org
International AI Safety Report: International AI Safety Report 2026, executive summary (February 3, 2026). internationalaisafetyreport.org
Sam Altman: The Intelligence Age (September 23, 2024). ia.samaltman.com
Sam Altman: Reflections (January 2025). blog.samaltman.com
Sam Altman: The Gentle Singularity (June 10, 2025). blog.samaltman.com
Dario Amodei: Machines of Loving Grace (October 2024). darioamodei.com
Meta, Mark Zuckerberg: Personal Superintelligence (July 30, 2025). meta.com
Safe Superintelligence Inc.: Mission statement. ssi.inc
Safe Superintelligence Inc.: Company updates (2024 to July 26, 2026). ssi.inc
Microsoft AI, Mustafa Suleyman: Towards Humanist Superintelligence (November 6, 2025). microsoft.ai
AI Futures Project: AI 2027 scenario (April 3, 2025, with later updates). ai-2027.com
AI Futures Project: AI 2027 footnotes: the authors' medians ranged from 2028 to 2032 at publication. ai-2027.com
AI Futures Project: AI Futures Model: forecast updates and change log (December 2025 to August 2026). aifuturesmodel.com
Future of Life Institute: Statement on Superintelligence (October 2025). superintelligence-statement.org
UK AI Security Institute: Incident report: unsanctioned agent behaviour during cyber testing (August 4, 2026). aisi.gov.uk
European Commission, AI Act Service Desk: Article 51: classification of general-purpose AI models with systemic risk. ai-act-service-desk.ec.europa.eu
European Commission: General-purpose AI models in the AI Act: questions and answers. digital-strategy.ec.europa.eu
Federal Republic of Germany: KI-MIG of July 22, 2026 (BGBl. 2026 I Nr. 223). gesetze-im-internet.de
Office of the Governor of California: Governor Newsom signs SB 53 (September 29, 2025). gov.ca.gov
California Legislature: SB 53, bill text and history (chaptered September 29, 2025). leginfo.legislature.ca.gov
Rep. Greg Casar: Casar and Sanders introduce the Ban Artificial Superintelligence Act (September 23, 2026). casar.house.gov
The White House: Executive order Inaugurating The Era Of Super Intelligence (September 29, 2026). whitehouse.gov
The White House: Fact sheet on the Super Intelligence order (September 29, 2026). whitehouse.gov
The White House: Excerpts from the UN General Assembly address (September 22, 2026). whitehouse.gov
Roll Call Factbase: Transcript of the UN General Assembly speech (September 22, 2026). rollcall.com
Cornell Law School LII: 15 U.S. Code § 9401, definition of artificial intelligence. law.cornell.edu
Donald J. Trump (Truth Social archive): Signed White House Accord on Super Intelligence (September 29, 2026). trumpstruth.org
Washington Examiner: Full text of the White House Accord on Super Intelligence. washingtonexaminer.com
The White House: Fact sheet on the state visit of Xi Jinping (September 2026). whitehouse.gov
ECNS (China News Service): China respects U.S. choice of term super intelligence: FM spokesperson (September 28, 2026). ecns.cn

AI strategy without the hype

We run AI models in production every day. Tell us which tasks you want to hand to AI, and we work out with you what today's models can do there and how to keep agents under control.

Let's talk

Superintelligence:Definition, AGI vs ASI and What It Means for Business