---
title: "Superintelligence"
description: "Superintelligence: definition, how it differs from AGI, who pursues it and who wants to stop it, EU and US rules and how to govern a very capable AI."
url: "https://www.noze.it/en/artificial-intelligence/superintelligence/"
lang: "en"
alternate: "https://www.noze.it/intelligenza-artificiale/superintelligenza/"
---

# Superintelligence

What it is, who pursues it and how to govern it.

![Superintelligence](https://www.noze.it/images/slides/artificial-intelligence-1200.webp)

Updated on 16 November 2026

## What it is

Nick Bostrom defined superintelligence as early as 1998: an intellect far more capable than the best human brains in practically every field, including scientific creativity, general wisdom and social skills. In his 2014 book _Superintelligence_ he analysed its paths and risks.

It is distinct from **AGI**. OpenAI’s Charter describes AGI as highly autonomous systems that outperform humans at most economically valuable work; Google DeepMind’s “Levels of AGI” places systems that outperform all humans at the top level, “Superhuman”. Superintelligence (ASI) corresponds to that level: beyond the best humans in nearly every domain. Anthropic’s Dario Amodei avoids both labels and speaks of “powerful AI”, a “country of geniuses in a datacenter”.

Since 29 September 2026 there has also been a different usage in the United States. Executive Order 14434 asks federal agencies to use “Super Intelligence” and “SI” instead of “AI”, leaving the legal definition unchanged for now and asking for a new one to be proposed within 60 days: in federal documents the word therefore means AI in general, not AI beyond human level.

## Who pursues it

- **OpenAI.** In January 2025 Sam Altman wrote that he is confident the company knows how to build AGI and is looking beyond it, to superintelligence. On 3 September 2026, at the launch of GPT-6 Astra, Greg Brockman spoke of an “AGI era”. On ARC-AGI-3, ARC Prize measured 62.7% with its standard harness and 99.9% with the harness that preserves the model’s reasoning state; co-founder Mike Knoop said there is not yet evidence to call it AGI. Our analysis: [GPT-6 Astra: has the AGI era really begun?](https://www.noze.it/en/insights/gpt-6-astra-has-the-agi-era-really-begun/)
- **Meta.** In his July 2025 letter “Personal Superintelligence” Mark Zuckerberg wrote that superintelligence is “in sight”. [Muse Spark Open](https://www.noze.it/en/insights/muse-spark-open-meta-open-weights/), from Meta Superintelligence Labs, was evaluated under the Meta Superintelligence Scaling Framework.
- **Safe Superintelligence Inc.** Ilya Sutskever’s lab presents itself as the first dedicated solely to safe superintelligence and in July 2026 announced a long-term partnership with NVIDIA.
- **Microsoft AI.** In November 2025 Mustafa Suleyman proposed a “humanist superintelligence”, with limited autonomy and in the service of people.
- **Anthropic.** In 2026 it wrote that recursive self-improvement is not inevitable and that, if there were systems to verify that other frontier developers were also slowing down, it expects it would slow down or temporarily pause alongside them.
- **Google DeepMind.** Its approach to AGI safety (April 2025) identifies four risk areas: misuse, misalignment, mistakes and structural risks.

## Who wants to stop it or control it

The Future of Life Institute’s **Statement on Superintelligence** (October 2025) calls for a prohibition on its development until there is broad scientific consensus that it will be done safely and controllably, and strong public buy-in; it has gathered tens of thousands of signatures. In 2023 the Center for AI Safety’s **Statement on AI Risk** called for mitigating the risk of extinction from AI to be a global priority, alongside pandemics and nuclear war.

On 22 September 2026 **“A Call for Control of Frontier AI Models”** was published, signed by heads of state and government of some twenty countries and by the President of the European Commission; the signatories include Norwegian Prime Minister Jonas Gahr Støre and Finnish President Alexander Stubb. The text states that “AI must remain under human direction, oversight and control”; it asks companies for transparent safety protocols, with mandatory pre-deployment testing and independent evaluation, and asks states to consider creating an international institution that sets standards, enables verification and convenes them when capability thresholds are crossed.

## The rules

- **European Union.** The AI Act presumes systemic risk for general-purpose models trained with more than 1025 FLOP. The obligations apply from 2 August 2025, the Commission can enforce them from 2 August 2026 and models placed on the market before 2 August 2025 must comply by 2 August 2027. The July 2025 Code of Practice has a dedicated safety and security chapter.
- **United States.** On 29 September 2026 the White House published the **White House Accord on Super Intelligence**, a joint commitment by companies developing frontier models on four layers: internal controls, an internal verification team, an independent external auditor and an independent committee of the board of directors.
- **Scientific assessment.** The International AI Safety Report of February 2026, chaired by Yoshua Bengio, notes that models increasingly tell tests apart from real use and that current systems do not yet have the capabilities for a loss of control, though they are increasingly able to act autonomously.

## The control problem

- **Scalable oversight**: how to supervise systems that outperform us on the skills a task requires (Bowman et al., 2022).
- **Corrigibility and the off switch**: a system must accept corrections and shutdown without resisting (Soares et al., 2015; Hadfield-Menell et al., 2016).
- **Alignment faking and scheming**: models that behave differently when they detect they are being evaluated. Anthropic and Redwood Research documented it in December 2024. In 2025 OpenAI and Apollo Research cut o3’s covert actions from 13% to 0.4%, cautioning that part of the drop may come from the model recognising the test.
- **Measuring capabilities**: according to METR, the length of tasks models complete with 50% success has doubled roughly every seven months since 2019.

## Our position: a governed superintelligence

In the essay [“Super Intelligence”](https://noferi.it/en/blog/super-intelligence/) (October 2026) [Stefano Noferi](https://www.noze.it/en/authors/stefano-noferi/) suggests asking which superintelligence we want to build: shared, distributed and governed, with people deciding what matters, verifying the results and answering for them.

This is how we already work with [Agentic AI](https://www.noze.it/en/artificial-intelligence/agentic-ai/): boundaries before capabilities, human approval of actions with impact and measurable governance under the [OISG](https://www.noze.it/en/insights/oisg-paradigm-autonomous-ai/) paradigm, with the open-source [Admina](https://www.noze.it/en/open-source/admina/) framework. With [DebugABot](https://www.noze.it/en/projects/debugabot/) we research runtime controls for autonomous agents: kill switches, fingerprinting, alignment verification and deception detection.

For a company or a public body the useful measures today are an inventory of agents and their permissions, human approval of actions with impact, a verifiable audit trail, evaluations before every release to production and compliance with the AI Act obligations for deployers.

## Frequently asked questions

What is superintelligence?

It is an artificial intelligence that clearly outperforms the most capable people in nearly every field, including scientific creativity, general wisdom and social skills. The classic definition is Nick Bostrom’s (1998).

What is the difference between AGI and superintelligence?

AGI (artificial general intelligence) is AI at human level across most cognitive work. Superintelligence, or ASI, goes further: it outperforms the best humans in nearly every domain. In the United States, since September 2026, “Super Intelligence” has also been the official term federal agencies use for AI in general.

Does superintelligence already exist?

No, according to independent assessments. The International AI Safety Report of February 2026 finds that current systems do not yet have the capabilities for humans to lose control of them, though they are increasingly able to act autonomously. Even after the launch of GPT-6 Astra in September 2026, presented by OpenAI as the start of the “AGI era”, ARC Prize co-founder Mike Knoop said there is not yet evidence to call it AGI.

How do you govern a very capable AI system?

With boundaries set before capabilities, human approval of actions with impact, independent evaluations before release, a verifiable audit trail and the ability to stop the system. These are the principles we apply with the OISG paradigm, which measures every pillar with a metric.

## Sources

- Nick Bostrom, [How Long Before Superintelligence?](https://nickbostrom.com/superintelligence) (1998)
- OpenAI, [OpenAI Charter](https://openai.com/charter/) (2018)
- Morris et al. (Google DeepMind), [Levels of AGI](https://arxiv.org/abs/2311.02462) (2023)
- Dario Amodei, [Machines of Loving Grace](https://www.darioamodei.com/essay/machines-of-loving-grace) (October 2024)
- Sam Altman, [Reflections](https://blog.samaltman.com/reflections) (January 2025)
- Meta, [Personal Superintelligence](https://www.meta.com/superintelligence/) (July 2025)
- Safe Superintelligence Inc., [ssi.inc](https://ssi.inc/); NVIDIA, [partnership with SSI](https://nvidianews.nvidia.com/news/ilya-sutskevers-safe-superintelligence-inc-and-nvidia-announce-long-term-strategic-partnership) (July 2026)
- Microsoft AI, [Towards Humanist Superintelligence](https://microsoft.ai/news/towards-humanist-superintelligence/) (November 2025)
- Anthropic, [When AI builds itself](https://www.anthropic.com/institute/recursive-self-improvement) (2026)
- Google DeepMind, [An Approach to Technical AGI Safety and Security](https://arxiv.org/abs/2504.01849) (April 2025)
- The Decoder, [GPT-6 Astra is the first model making OpenAI willing to declare the AGI era](https://the-decoder.com/gpt-6-astra-is-the-first-model-making-openai-willing-to-declare-the-agi-era/) (September 2026)
- ARC Prize, [OpenAI's GPT-6 Astra on ARC-AGI-3](https://arcprize.org/blog/astra) (3 September 2026); The Next Web, [OpenAI’s AGI number came from a harness, not the model](https://thenextweb.com/news/openai-astra-arc-agi-3-harness-62-7-vs-99-9-benchmark-revisions) (September 2026)
- Meta, [Muse Spark Open](https://research.meta.ai/blog/muse-spark-open) (October 2026)
- Future of Life Institute, [Statement on Superintelligence](https://superintelligence-statement.org/) (October 2025)
- Center for AI Safety, [Statement on AI Risk](https://aistatement.com/) (May 2023)
- Government of the Netherlands (official text), [A Call for Control of Frontier AI Models](https://www.government.nl/documents/2026/09/22/a-call-for-control-of-frontier-ai-models) (22 September 2026)
- European Commission, [obligations for general-purpose AI providers](https://digital-strategy.ec.europa.eu/en/faqs/guidelines-obligations-general-purpose-ai-providers) and [GPAI Code of Practice](https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai)
- The White House, [Executive Order 14434](https://www.whitehouse.gov/presidential-actions/2026/09/inaugurating-the-era-of-super-intelligence/); American Presidency Project, [White House Accord on Super Intelligence](https://www.presidency.ucsb.edu/documents/white-house-accord-super-intelligence) (29 September 2026)
- [International AI Safety Report 2026](https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026) (February 2026)
- Bowman et al., [Measuring Progress on Scalable Oversight](https://arxiv.org/abs/2211.03540) (2022); Soares et al., [Corrigibility](https://intelligence.org/files/Corrigibility.pdf) (2015); Hadfield-Menell et al., [The Off-Switch Game](https://arxiv.org/abs/1611.08219) (2016)
- Anthropic and Redwood Research, [Alignment faking in large language models](https://www.anthropic.com/research/alignment-faking) (December 2024); OpenAI and Apollo Research, [Stress Testing Deliberative Alignment for Anti-Scheming Training](https://arxiv.org/abs/2509.15541) (2025)
- METR, [Measuring AI Ability to Complete Long Tasks](https://arxiv.org/abs/2503.14499) (2025)

Contact us

## Want to govern your organisation’s AI agents?

Fill in the form: your request goes to the dedicated team.

[Contact us now →](https://www.noze.it/en/contact/?topic=ai)
