The definition, and where it comes from
Bostrom defines a superintelligence as "any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest" [1]. Three details in how he frames it matter.
- It is about general ability, not one skill. Bostrom says a chess program is not a superintelligence because it is only smart within chess [1].
- It does not say how the intellect is built, and it does not require consciousness [1].
- He allows narrower uses when they are stated explicitly, such as an "engineering superintelligence" that vastly outperforms the best human minds in engineering [1].
In an earlier short essay he put it as "a general intelligence that vastly outperforms the best human brains in every significant cognitive domain" [4].
Other writers use stricter or looser wording. Google DeepMind researchers, for example, define the top level of their AGI scale ("ASI") as a general system that outperforms 100% of humans on a wide range of tasks [3]. IBM describes artificial superintelligence as a hypothetical software-based AI system with an intellectual scope beyond human intelligence [2]. These are close in spirit but not identical, and there is no single standard definition.
SI, ASI, AGI and related terms
People use these words loosely. This table shows what each usually means and who uses it. The companion page, AI vs AGI vs superintelligence, goes deeper on the differences.
| Term | What it usually means | Source |
|---|---|---|
| Superintelligence | An intellect far beyond humans in virtually all domains of interest | Bostrom [1] |
| ASI (artificial superintelligence) | The same idea, stressing that it would be an artificial system | IBM [2]; DeepMind's "Level 5" general AI [3] |
| SI | Shorthand for superintelligence in AI writing. Also the name of the International System of Units, and since September 2026 a US executive branch naming term. See below and SI vs AI. | [5] [6] |
| AGI (artificial general intelligence) | An AI system roughly as capable as a human across most cognitive tasks. Definitions vary widely. | [3] |
| Narrow AI | AI built or used for specific tasks, even if it beats every human at that task | [3] [2] |
| "Humanist superintelligence" | Microsoft AI's term for very advanced, problem-oriented, tightly controlled systems that tend to be domain specific | [7] |
| "Personal superintelligence" | Meta's term for AI that knows a person's goals and helps them reach them | [8] |
The last two rows are company usages. They show the word being stretched, not a shared technical meaning, so when you read "superintelligence" in a product announcement, check which sense is meant.
Is it one thing? Speed, collective and quality
Bostrom separates three ways something could be superintelligent: it thinks much faster than a human (speed), it is a system of many smaller intellects whose combined performance far exceeds any current cognitive system (collective), or it is simply much better at thinking at the same speed (quality) [1].
The collective form is the one most relevant to tools that combine several AI systems. It is a category in a philosophical taxonomy. It is not evidence that combining today's models produces superintelligence, and we do not claim that. More on this in multi-model AI.
Proposed paths to superintelligence
Bostrom's book surveys several routes: artificial intelligence, whole brain emulation, enhancing biological cognition, brain-computer interfaces, and networks and organizations [1]. His 2009 essay highlights two ideas: human-level machine intelligence followed by a much quicker step to superintelligence, and the "singularity hypothesis", in which a "seed AI" improves itself over repeated cycles, producing what he calls an intelligence explosion [4].
He also writes that biological human brains might one day cease to be the predominant nexus of intelligence, while arguing that improving individual and collective human intelligence is worth pursuing in the meantime [4].
Today the public paths are company strategies. Safe Superintelligence Inc. says building safe superintelligence is its one goal and product [9]. Meta states a vision of "personal superintelligence" [8]. Microsoft AI announced a team working toward "humanist superintelligence" [7]. OpenAI's leaders have written about the governance of future systems "dramatically more capable than even AGI" [10]. These are statements of intent, not demonstrations.
What organizations have said, with dates
| Who | Date | Position in their own words or a close paraphrase |
|---|---|---|
| OpenAI (Altman, Brockman, Sutskever) | 22 May 2023 | Superintelligence will need special treatment and coordination; they suggest something like an international authority for efforts above a capability or compute threshold, and say stopping its creation would be "unintuitively risky and difficult" [10] |
| Safe Superintelligence Inc. | undated home page | "Superintelligence is within reach"; building a safe one is the company's sole focus [9] |
| Meta (Zuckerberg) | 30 July 2025 | Aims to bring personal superintelligence to everyone; says it will raise novel safety concerns and calls for care about what is open sourced [8] |
| Microsoft AI (Suleyman) | 6 November 2025 | Working toward "humanist superintelligence": problem-oriented, domain-specific tendencies, kept within limits, with humans in control [7] |
| Future of Life Institute and signatories | 22 October 2025 | Calls for a prohibition on developing superintelligence until there is broad scientific consensus that it can be done safely and controllably and strong public buy-in [11] |
These positions disagree. OpenAI's 2023 piece argues that halting development would be hard and risky, while the 2025 statement asks for a prohibition until conditions are met. News coverage reported more than 800 signatories at launch [12]. Reading one side alone gives a lopsided picture.
Risks and the safety debate
Bostrom wrote that it is an open question whether the consequences would be for the better or the worse: the potential upside is enormous, but the downside includes existential risk, and outcomes might depend on the initial conditions we create [4]. The core worry is control: how to make sure a system much more capable than its designers pursues goals people actually want.
Survey data shows researchers are divided. In AI Impacts' 2023 survey of 2,778 researchers who had published in top AI venues, 68.3% thought good outcomes from superhuman AI are more likely than bad ones, yet between 38% and 51% (depending on how the question was asked) gave at least a 10% chance to outcomes as bad as human extinction [13]. That is a survey of opinions about a hypothetical, from one year, not a measurement.
Google DeepMind's researchers argue that risks differ by capability level and by how much autonomy a system has, and that the "Virtuoso" and "ASI" levels are where many concerns about extreme risk are most likely to emerge [3]. Microsoft's Suleyman argues for putting human control first, then accelerating [7].
Timelines are contested opinion, not fact
Nobody can say when, or whether, superintelligence will exist. What exists is a range of forecasts:
- Bostrom (2009): whether human-level AI takes a couple of decades, many decades or centuries is "probably not something that we are currently in a position to know" [4].
- AI Impacts' 2023 survey: aggregate forecasts gave a 10% chance by 2027 and 50% by 2047 that unaided machines outperform humans in every possible task. That is a different and narrower claim than superintelligence, and the same survey put a 50% chance of all human occupations being fully automatable as late as 2116 [13].
- OpenAI's leaders (2023): "conceivable" that within ten years AI systems will exceed expert skill level in most domains [10].
- Meta (2025): the rest of this decade seems likely to be the decisive period for the path this technology takes [8].
These come from different people with different incentives and definitions. The 2023 survey also moved by 13 years against the 2022 one [13], which is a reminder of how unstable such forecasts are. We do not offer our own forecast.
"SI" has more than one meaning
In this page's sense, SI is shorthand for superintelligence. But the letters are also the symbol for the International System of Units, the modern metric system maintained by the BIPM [5].
And in September 2026 a US executive order directed the executive branch to use "Super Intelligence" and "SI" in place of "Artificial Intelligence" and "AI" in many official documents, while defining the new terms as the same technologies the existing legal definition of AI covers [6]. That is a naming change inside government. It does not mean the systems it covers are superintelligent in Bostrom's sense. The full details are on SI vs AI.