---
id: ASI-DEF-001
title: "Defining and Measuring Superintelligence"
alternative_names:
  - "Definitions of Superintelligence"
  - "Superhuman Intelligence: Criteria and Measurement"
category: "Advanced Intelligence → Superintelligence"
volume: VI
chapter: 31
entry_type: "Standard Article"
level: 3
status: "Draft"
last_reviewed: 2026-10-15
version: "0.1.0"
prerequisites:
  - "ASI-001"
  - "AGI-DEF-001"
related_entries:
  - "ASI-EXP-001"
  - "ASI-PATH-001"
  - "AGI-TEST-001"
  - "FOUND-002"
---

# Defining and Measuring Superintelligence

## 1. Concise Definition
**Superintelligence** is a hypothetical level of intelligence that greatly exceeds that of humans. The most cited definition is Bostrom's: "an intellect that is much smarter than the best human brains in practically every field" (Bostrom, 1998). No system is known to meet any proposed definition of general superintelligence, and how such a property would be measured is itself unresolved.

## 2. Overview & Significance
[ASI-001](../vol06_asi/ASI-001_artificial_superintelligence.md) introduces artificial superintelligence (ASI) and its implications. This article addresses a prior question: what the word means and how one could tell that a system satisfies it. The question is harder than the corresponding one for AGI ([AGI-DEF-001](../vol05_agi/AGI-DEF-001_definitions_and_operational_criteria_for_agi.md)), for three reasons.

1. Human-relative scales have a ceiling. Percentile-of-humans measures cannot distinguish a system that outperforms the best human slightly from one that outperforms all humans by a large margin.
2. Verification becomes difficult. If a system's outputs exceed the ability of human evaluators to check them, the evaluator cannot confirm correctness directly (see [SAFE-OVERSIGHT-001](../vol08_safety_governance/SAFE-OVERSIGHT-001_scalable_oversight_and_corrigibility.md)).
3. Superhuman performance already exists in narrow domains, so "superhuman" without a scope is ambiguous.

The distinction between narrow and general superhuman performance is central. Systems that exceed the best humans at specific tasks, such as Go, chess and shogi (Silver et al., 2018) or protein structure prediction (Jumper et al., 2021), are not usually called superintelligent in the sense discussed here.

## 3. Historical Development
Good (1966) described an "ultraintelligent machine" as one that could far surpass the intellectual activities of any person, however clever, and argued that because designing machines is an intellectual activity, such a machine could design better ones. (The paper is often dated 1965; the volume of *Advances in Computers* in which it appears is dated 1966 by its publisher.) This is the origin of the intelligence-explosion hypothesis treated in [ASI-EXP-001](../vol06_asi/ASI-EXP-001_intelligence_explosion.md), which is Hypothetical and Contested; the definitional point is that Good's definition is comparative and unbounded.

Bostrom (1998) gave the definition quoted above and discussed when hardware might match estimates of brain processing power. Chalmers (2010) analysed the "singularity" philosophically and distinguished levels beyond human intelligence. [VERIFICATION NEEDED: the exact labels and definitions of Chalmers's levels (AI, AI+, AI++) were not checked against the paper]. Bostrom (2014) classified superintelligence by the way it exceeds humans, into speed, collective and quality forms. Morris et al. (2024) placed "Superhuman" as the top performance level of a general taxonomy, defined as outperforming 100% of humans.

Formal approaches treat intelligence as a measurable quantity. Legg and Hutter (2007) defined universal intelligence, in which an agent's score is its expected performance over computable environments, weighted by simplicity. On this definition superintelligence corresponds to a score exceeding that of any human, but the definition is not computable and the weighting is a theoretical choice (see [AGI-DEF-001](../vol05_agi/AGI-DEF-001_definitions_and_operational_criteria_for_agi.md)).

## 4. Technical / Conceptual Explanation

### 4.1 Dimensions of "much smarter"
Definitions differ in three dimensions that should be specified together.

- **Scope**: "practically every field" (Bostrom, 1998) versus a specified set of tasks (Morris et al., 2024, whose matrix pairs performance with generality).
- **Margin**: any outperformance of the best human, or a large factor. Bostrom (1998) says "much smarter"; no numeric margin is fixed.
- **Reference**: the best individual human, the best human team, or humanity as a whole with its tools and institutions.

The reference matters. A system might outperform every individual human at every cognitive task while remaining weaker than human civilisation, which can pool many people, specialisations and tools. Bostrom (2014) addresses this through the distinction among forms below.

### 4.2 Forms of superintelligence (Bostrom, 2014)
- **Speed superintelligence**: does what a human mind does, but faster.
- **Collective superintelligence**: a system composed of many smaller intellects whose aggregate performance greatly exceeds that of any current group.
- **Quality superintelligence**: a system that is at least as fast as a human mind and substantively better in the quality of its thinking.

These are conceptual categories, and whether any form is achievable is Hypothetical.

### 4.3 Measuring performance above the human range
Several approaches have been suggested, each with limitations.

1. **Task-specific rating scales.** In games, competitive ratings (Elo-style) can place a system above the best humans. They require a population of opponents, and a rating for a system far above the population is poorly determined.
2. **Difficulty-scaled benchmarks.** Tasks with ground-truth verification (formal proofs, code with tests, physical experiments) allow measurement beyond human ability if the problems can be generated or verified automatically. Many real-world tasks lack cheap verification.
3. **Human-relative time metrics.** The time-horizon metric of Kwa et al. (2025) measures the length of tasks, in human-expert time, that a model completes at a given success rate. The approach is defined for software tasks and is bounded by the length of tasks for which human baselines can be obtained.
4. **Economic or scientific outcomes.** Measures such as discoveries or output value avoid dependence on test design but are slow, noisy and confounded.
5. **Proxy profiles.** Multi-domain schemes such as Hendrycks et al. (2025) score against a "well-educated adult" and do not extend naturally above that level.

### 4.4 The verification problem
When a system proposes a solution that humans cannot evaluate, the claim "the system is superhuman" cannot be checked by comparison with human answers. Burns et al. (2023) study weak-to-strong generalisation: whether a strong model supervised by a weaker one can recover capability beyond its supervisor. They treat this as an analogy for humans supervising superhuman models, and it is a research setting, not evidence about actual superhuman systems.

## 5. Types, Components & Classifications

| Classification | Categories | Source |
| :--- | :--- | :--- |
| By generality | Narrow superhuman; general superhuman | Silver et al. (2018); Bostrom (1998) |
| By form | Speed; collective; quality | Bostrom (2014) |
| By performance level | Level 5 "Superhuman" in a general matrix | Morris et al. (2024) |
| By formal score | Exceeds all humans on universal intelligence | Legg & Hutter (2007) |
| By reference class | Best individual; best team; civilisation | Section 4.1 |

Terms related to superintelligence include "ultraintelligent machine" (Good, 1966), "strong superintelligence" and "the singularity". The last denotes a hypothesised event rather than a type of entity.

## 6. Applications & Concrete Examples
- **Narrow superhuman systems.** Self-play systems for board games achieved performance above the best human players (Silver et al., 2018). AlphaFold predicted protein structures with high accuracy in the CASP14 assessment (Jumper et al., 2021). These are laboratory or scientific results in specific domains, and neither is a general system.
- **Policy use of the term.** Statements by organisations and public figures use "superintelligence" with varying meanings. Because definitions vary, such statements need to be read with attention to the definition used. Current statements are not catalogued in this stable article.
- **Evaluation design.** Weak-to-strong generalisation experiments (Burns et al., 2023) and similar oversight research use defined proxies for superhuman capability, such as a weak supervisor and a stronger student model.

## 7. Advantages, Limitations & Unresolved Questions
**Value of precise definitions.** Clear definitions allow safety and governance discussions to specify which capability levels trigger concern, and prevent conflating narrow superhuman tools with general ones.

**Limitations.**
- The standard definitions are qualitative; "much smarter" has no agreed margin.
- Measures rely on human reference points that stop discriminating at the top.
- The forms in Bostrom (2014) are categories in an argument, not measured quantities.
- Intelligence may not be a single scalar. If it is multidimensional, "superior in practically every field" would require a profile comparison, and incomparable profiles are possible.

**Open questions.**
1. Is there a coherent scalar notion of intelligence above the human range? This is Contested.
2. Can superhuman general ability be verified without trusting the system?
3. Does superintelligence require qualitatively new capabilities, or is it quantitative extension of human-like ones?
4. Where does a collective of human-level systems operating at scale fall: AGI or superintelligence? Under Bostrom's collective form, potentially the latter.

## 8. Current State of Knowledge
As of 2026-10-15, no source reviewed shows a system that meets a general definition of superintelligence, and all substantive claims about superintelligence are Hypothetical. Narrow superhuman performance is documented for particular tasks (Silver et al., 2018; Jumper et al., 2021). Definitions of superintelligence are not standardised; the definitions of Bostrom (1998, 2014), Good (1966), Morris et al. (2024) and Legg and Hutter (2007) coexist and answer different questions. Measurement methods above the human range are an active research topic without an accepted standard. Current model assessments are collected in the Living Reference (`MODEL-001`, `FACT-001`).

## 9. Future Trajectory & Strategic Outlook
This section is descriptive and registers no forecast.

- **Near term (0 to 3 years).** Further narrow superhuman results in verifiable domains are likely to continue to be reported, along with research on supervision of models that exceed their evaluators (Burns et al., 2023).
- **Medium term (3 to 10 years).** Development of measurement frameworks that function above the human range, for example based on automatically verifiable tasks, would clarify the concept. The sources reviewed do not show that such frameworks exist now.
- **Long term (10+ years).** Whether superintelligence in Bostrom's sense is achievable is Hypothetical. Dated forecasts of ASI belong in the Forecast Register (`FCST-REG-001`) and use the format in the evidence framework.

## 10. Related Entries & Knowledge Graph
- **Foundational Concepts**:
  - [[ASI-001] Artificial Superintelligence](../vol06_asi/ASI-001_artificial_superintelligence.md): Overview of ASI and its implications.
  - [[AGI-DEF-001] Definitions and Operational Criteria for AGI](../vol05_agi/AGI-DEF-001_definitions_and_operational_criteria_for_agi.md): The corresponding problem at the human level.
  - [[FOUND-002] Definitions of Intelligence](../vol01_foundations/FOUND-002_definitions_of_intelligence.md): General definitions of intelligence.
- **Related Methods & Technologies**:
  - [[AGI-TEST-001] Testing Claims of General Intelligence](../vol05_agi/AGI-TEST-001_testing_claims_of_general_intelligence.md): Test validity issues that grow above the human range.
  - [[SAFE-OVERSIGHT-001] Scalable Oversight and Corrigibility](../vol08_safety_governance/SAFE-OVERSIGHT-001_scalable_oversight_and_corrigibility.md): Supervising systems that exceed their evaluators.
- **Advanced Implications & Future Trajectories**:
  - [[ASI-EXP-001] The Intelligence Explosion Hypothesis](../vol06_asi/ASI-EXP-001_intelligence_explosion.md): Hypothesised route from human-level to superhuman.
  - [[ASI-PATH-001] Pathways to Superintelligence](../vol06_asi/ASI-PATH-001_pathways_to_superintelligence.md): Proposed routes to superintelligence.
  - [[BIO-BOSTROM-001] Nick Bostrom](../vol10_people_organizations/BIO-BOSTROM-001_nick_bostrom.md): Author of the principal definitions.

## 11. References & Foundational Literature
1. Bostrom, N. (1998). *How Long Before Superintelligence?* International Journal of Futures Studies, 2. Reprinted in Linguistic and Philosophical Investigations, 5(1), 11–30 (2006). https://nickbostrom.com/superintelligence. [Source of the definition quoted in Section 1.]
2. Bostrom, N. (2014). *Superintelligence: Paths, Dangers, Strategies*. Oxford University Press. ISBN: 978-0199678112. [Speed, collective and quality forms of superintelligence.]
3. Burns, C., Izmailov, P., Kirchner, J. H., Baker, B., Gao, L., Aschenbrenner, L., Chen, Y., Ecoffet, A., Joglekar, M., Leike, J., Sutskever, I., & Wu, J. (2023). *Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision*. arXiv preprint. arXiv:2312.09390. [Experimental analogy for supervising models that exceed their supervisors.]
4. Chalmers, D. J. (2010). *The Singularity: A Philosophical Analysis*. Journal of Consciousness Studies, 17(9–10), 7–65. [Philosophical analysis of the singularity and levels beyond human intelligence. [VERIFICATION NEEDED: issue number differs between sources, and the level labels were not checked in the paper]]
5. Good, I. J. (1966). *Speculations Concerning the First Ultraintelligent Machine*. Advances in Computers, 6, 31–88. https://doi.org/10.1016/S0065-2458(08)60418-0. [Origin of the ultraintelligent machine concept.]
6. Hendrycks, D., Song, D., Szegedy, C., Lee, H., Gal, Y., Brynjolfsson, E., et al. (2025). *A Definition of AGI*. arXiv preprint. arXiv:2510.18212. [Profile scoring against a well-educated adult; illustrates the human-relative ceiling.]
7. Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., et al. (2021). *Highly accurate protein structure prediction with AlphaFold*. Nature, 596(7873), 583–589. https://doi.org/10.1038/s41586-021-03819-2. [Narrow superhuman-class result in protein structure prediction.]
8. Kwa, T., West, B., Becker, J., Deng, A., Garcia, K., Hasin, M., et al. (2025). *Measuring AI Ability to Complete Long Software Tasks*. NeurIPS 2025; arXiv:2503.14499. [Human-relative time-horizon metric.]
9. Legg, S., & Hutter, M. (2007). *Universal Intelligence: A Definition of Machine Intelligence*. Minds and Machines, 17(4), 391–444. https://doi.org/10.1007/s11023-007-9079-x. [Formal definition from which an above-human score can be defined.]
10. Morris, M. R., Sohl-Dickstein, J., Fiedel, N., Warkentin, T., Dafoe, A., Faust, A., Farabet, C., & Legg, S. (2024). *Levels of AGI for Operationalizing Progress on the Path to AGI*. Proceedings of ICML 2024; arXiv:2311.02462. [Superhuman level defined as outperforming 100% of humans.]
11. Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A., Lanctot, M., Sifre, L., Kumaran, D., Graepel, T., Lillicrap, T., Simonyan, K., & Hassabis, D. (2018). *A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play*. Science, 362(6419), 1140–1144. https://doi.org/10.1126/science.aar6404. [Narrow superhuman performance in board games.]

## 12. Evidential Classification Matrix
| Section / Claim | Evidence Category | Supporting Source | Confidence / Status |
| :--- | :--- | :--- | :--- |
| Narrow superhuman performance exists in some games and prediction tasks | Well Supported | Silver et al. (2018); Jumper et al. (2021) | High |
| A general system meeting Bostrom's definition exists | Not supported by sources reviewed | n/a | n/a |
| Speed, collective and quality forms are achievable | Hypothetical | Bostrom (2014) | Low |
| Intelligence above the human range is a coherent scalar | Contested | Legg & Hutter (2007); Bostrom (2014) | Active debate |
| Ultraintelligent machine would design better machines (intelligence explosion) | Hypothetical / Contested | Good (1966) | Low; see ASI-EXP-001 |
| Weak-to-strong generalisation experiments are a proxy for supervising superhuman systems | Preliminary | Burns et al. (2023) | Moderate |
| Measurement above the human range lacks an accepted standard | Well Supported (no counter-example in the sources reviewed) | Kwa et al. (2025); Hendrycks et al. (2025) | Moderate |

## 13. Revision History
- **v0.1.0** (2026-10-15): Initial draft. Citations checked against arXiv/Crossref/publisher pages where marked; not independently fact-checked.