---
id: FOUND-002
title: "Definitions of Intelligence"
alternative_names:
  - "Intelligence"
  - "Cognition"
  - "Cognitive Capability"
category: "Foundations → Cognition"
volume: I
chapter: 1
entry_type: "Standard Article"
level: 2
status: "Draft"
last_reviewed: 2026-10-15
version: "0.2.0"
prerequisites:
  - "FOUND-001"
related_entries:
  - "AGI-001"
  - "FOUND-005"
---

# Definitions of Intelligence

## 1. Concise Definition
**Intelligence** is a multifaceted property of biological and computational systems characterized by the capacity to acquire knowledge, perceive complex patterns, reason abstractly, and adaptively achieve complex goals in novel environments. Because intelligence manifests across diverse substrates—ranging from human cognition and animal behavior to collective swarms and artificial algorithms—there is no single universally accepted scientific definition, but rather a spectrum of operational, psychological, and computational frameworks used to measure and model cognitive competence (Legg & Hutter, 2007).

## 2. Overview & Significance
The effort to define intelligence is central to both cognitive science and artificial intelligence (AI). A precise definition establishes the criteria by which the success of AI systems is evaluated and determines the theoretical thresholds for achieving Artificial General Intelligence (AGI).

The difficulty in defining intelligence arises from its inherently unobservable nature; it is a latent trait inferred from external behavior and task performance. Different disciplines emphasize different dimensions:
1. **Psychology** emphasizes general cognitive factors ($g$), fluid reasoning, and linguistic competence.
2. **Evolutionary Biology** emphasizes adaptive fitness, morphological survival, and ecological problem-solving.
3. **Computer Science** emphasizes rational agency, computational complexity, and expected utility maximization within formalized state spaces.

Understanding these divergent definitions is crucial because optimizing an AI system for a narrow operational metric of intelligence often fails to yield the robust, generalized competence expected of human cognition.

## 3. Historical Development
Historically, human intelligence was treated as a singular, indivisible phenomenon heavily intertwined with consciousness and the soul. In the late 19th and early 20th centuries, Charles Spearman identified a general intelligence factor ($g$) that correlates performance across diverse cognitive tasks (Spearman, 1904), and Alfred Binet and Théodore Simon developed an early practical intelligence scale (1905). The "intelligence quotient" is usually credited to William Stern (1912) [VERIFICATION NEEDED: primary sources for Binet–Simon and Stern attributions].

With the advent of cybernetics in the 1940s, cybernetics researchers, including Norbert Wiener, framed goal-directed behavior as a property of information-processing systems rather than of biological tissue alone [VERIFICATION NEEDED: source for this characterization of Wiener and von Neumann]. In 1950, Alan Turing proposed replacing the question of whether machines think with a behavioral test, the Imitation Game (Turing Test), in which a machine's conversational responses are compared with a human's (Turing, 1950).

In the late 20th century, psychologists proposed multi-dimensional models, such as Howard Gardner's theory of multiple intelligences (Gardner, 1983), arguing against a monolithic $g$-factor. Conversely, mathematical computer scientists sought absolute universality. In 2007, Shane Legg and Marcus Hutter proposed a formal definition of universal machine intelligence based on algorithmic probability (Legg & Hutter, 2007).

## 4. Technical / Conceptual Explanation

### 4.1 Psychometric and Human-Centric Definitions
In psychology, intelligence is typically conceptualized as an aggregate capacity. A 1994 statement first published in the *Wall Street Journal* and signed by 52 of the 131 researchers invited to sign it (Gottfredson, 1997) opens its definition: *"Intelligence is a very general mental capability that, among other things, involves the ability to reason, plan,…"*, and goes on to list abilities such as abstract thinking, comprehension of complex ideas and learning from experience [VERIFICATION NEEDED: full wording to be checked against Gottfredson (1997)]. It was a signatory statement drafted in the context of the debate over *The Bell Curve*, not a consensus of the field.
This definition inherently anchors intelligence to the specific cognitive profile of *Homo sapiens*, making it problematic when applied to synthetic or alien architectures.

### 4.2 Behavioral and Task-Based Definitions
Turing's operational approach substituted a behavioral comparison for the question of whether machines think. Under this approach, if a system's observable behavior on a sufficiently broad set of tasks cannot be distinguished from a known intelligent entity (a human), it is deemed intelligent. Modern AI benchmarking (e.g., MMLU, Hendrycks et al., 2021; SWE-bench, Jimenez et al., 2024) relies heavily on this task-based framing, treating intelligence as aggregate accuracy over a battery of distinct tests.

### 4.3 Computational and Rational Agent Definitions
Artificial intelligence research frequently defines intelligence via rational agency. A rational agent maximizes its expected utility:
$$a^* = \arg\max_{a} \mathbb{E} [ U(s') \mid a, e ]$$
Intelligence, in this framework, is the computational efficiency and accuracy with which an agent selects actions that achieve its objectives given its available sensor data ($e$) and prior knowledge (Russell & Norvig, 2020).

### 4.4 Universal Machine Intelligence (Legg–Hutter)
Legg and Hutter (2007) provided a formalized, substrate-independent definition. They define intelligence $\Upsilon$ as an agent $\pi$'s expected capability to achieve goals across all possible computable environments $\mathcal{E}$, weighted inversely by the algorithmic complexity $K(\mu)$ of each environment:
$$\Upsilon(\pi) = \sum_{\mu \in \mathcal{E}} 2^{-K(\mu)} V_\mu^\pi$$
The measure is incomputable and depends on the choice of reference machine, which limits its practical use. It captures the intuition that a highly intelligent agent should succeed not only in simple, common environments but also in complex, novel ones.

## 5. Types, Components & Classifications

Intelligence can be structurally classified by its breadth and locus:
- **Narrow Intelligence (Competence)**: High performance optimized for a specific, bounded environment (e.g., a chess engine, a calculator). Does not adapt to out-of-domain tasks.
- **General Intelligence (Generality)**: The capacity to acquire new competencies across previously unseen, structurally diverse domains through transfer learning and abstraction.
- **Crystallized Intelligence**: The application of acquired knowledge, facts, and stored heuristics to solve problems.
- **Fluid Intelligence**: The ability to solve novel reasoning problems independent of previously acquired knowledge.
- **Collective Intelligence**: Cognitive capacities that emerge not from a single agent, but from the network topology and interactions of many simpler agents (e.g., ant colonies, decentralized human markets, ensemble machine learning models).

## 6. Applications & Concrete Examples
- **Psychometrics**: IQ testing and cognitive diagnostics used in clinical psychology and educational placement.
- **Biological Ethology**: Testing corvids and cephalopods for tool use, causal reasoning, and deferred gratification to study non-human biological intelligence.
- **AI Benchmarking**: Utilizing the abstraction of general intelligence to construct standardized suites (e.g., the Abstraction and Reasoning Corpus, Chollet, 2019), which tests abstract reasoning over novel visual puzzles and was designed to resist memorization.

## 7. Advantages, Limitations & Unresolved Questions

### Conceptual Limitations
- **Anthropocentrism**: Human-centric definitions penalize systems that solve problems using non-human heuristics, while task-based definitions (like the Turing Test) encourage deceptive mimicry over genuine comprehension.
- **The Measurement Problem**: Standardized AI benchmarks frequently suffer from Goodhart's Law; when a metric becomes an optimization target, it ceases to be a good measure. Systems can achieve high scores via dataset contamination or statistical correlation without possessing the underlying fluid reasoning the test was designed to measure.

### Decoupling Related Concepts
A recurring challenge is the conflation of intelligence with distinct philosophical and operational properties:
- **Consciousness / Sentience**: Subjective phenomenal experience. An AI can possess superintelligence without any subjective inner life (the "philosophical zombie" thought experiment; see [[FOUND-005]]).
- **Autonomy**: The ability to select self-determined goals. A highly intelligent system may be designed as an "oracle" that only answers questions, possessing zero autonomy.
- **Wisdom / Alignment**: Intelligence is purely the instrumental capacity to achieve a goal. It does not dictate that the goal itself is safe, ethical, or beneficial (the Orthogonality Thesis, a hypothesis rather than an established result; Bostrom, 2012).

### Unresolved Scientific Questions
1. *Is there a universal $g$-factor for artificial neural networks?* Do models that excel at natural language inherently develop better spatial or mathematical reasoning?
2. *Can intelligence exist without physical embodiment?* Does true causal understanding require physical interaction with a grounded environment, or is statistical sequence modeling sufficient?

## 8. Current State of Knowledge
Many treatments distinguish between instrumental capability (intelligence) and subjective experience (consciousness); this article does not claim a surveyed consensus. Debate continues on whether current Large Language Models possess "understanding." Formal definitions such as Legg and Hutter's (2007) have been influential theoretically, though the measure is incomputable and sensitive to the choice of reference machine, which limits practical use. However, empirical benchmarking remains heavily fragmented, with researchers actively seeking evaluation frameworks that resist rote memorization (Chollet, 2019).

## 9. Future Trajectory & Strategic Outlook
The following are plausible scenarios, not established facts, each registered as a forecast record.

- **Near-Term**: Plausible continued shift toward abstraction and reasoning benchmarks (e.g., ARC-AGI) alongside, or in place of, anthropomorphic evaluations such as the Turing Test (FCST-FOUND-002-001).
- **Medium-Term**: Plausible expansion of definitions of intelligence to cover multi-agent systems (FCST-FOUND-002-002).
- **Long-Term**: Possible, but uncertain, convergence on a unified theory of intelligence spanning neuroscience, algorithmic information theory and machine learning (FCST-FOUND-002-003).

| Forecast ID | Statement | Operational definition | Horizon | Assumptions | Confidence | Alternative scenarios | Review | Status |
| :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
| FCST-FOUND-002-001 | Abstraction-and-reasoning benchmarks are more widely reported than Turing-test results in frontier model evaluations. | Share of major frontier-model technical reports that include an abstraction/reasoning benchmark versus a Turing-test evaluation. | 2026–2029 | Benchmarks remain uncontaminated and widely adopted. | Medium (40–60%) | Other benchmark families displace both. | Annual, October 2027 | Active |
| FCST-FOUND-002-002 | Evaluation frameworks for multi-agent AI systems appear in a major benchmark suite or standard. | At least one widely used evaluation suite or standard explicitly scoring multi-agent behavior. | 2029–2036 | Multi-agent deployments become common. | Low–Medium (20–40%) | Evaluation stays at the single-model level. | Annual, October 2027 | Active |
| FCST-FOUND-002-003 | A unified theory of intelligence gains broad acceptance. | Adopted as a standard framework in major textbooks across psychology, neuroscience and AI. | After 2036 | Unspecified. | Low (<20%) | Definitions remain plural and domain-specific. | Annual, October 2027 | Active |

## 10. Related Entries & Knowledge Graph
- **Foundational Concepts**:
  - [[FOUND-001] Artificial Intelligence](./FOUND-001_artificial_intelligence.md): The overarching field concerned with synthesizing intelligence.
  - [[FOUND-005] Philosophy of Mind & The Turing Test](./FOUND-005_philosophy_of_mind.md): Deeper exploration of the mind-body problem and behavioral testing.
- **Advanced Implications & Future Trajectories**:
  - [[AGI-001] Artificial General Intelligence](../vol05_agi/AGI-001_artificial_general_intelligence.md): Systems exhibiting broad, human-equivalent fluid intelligence.

## 11. References & Foundational Literature
1. 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. [Comprehensive formalization of universal intelligence across computable environments].
2. Turing, A. M. (1950). *Computing Machinery and Intelligence*. Mind, 59(236), 433–460. https://doi.org/10.1093/mind/LIX.236.433. [The seminal paper introducing the Imitation Game].
3. Spearman, C. (1904). *General Intelligence, Objectively Determined and Measured*. American Journal of Psychology, 15(2), from p. 201 [end page not confirmed]. https://doi.org/10.2307/1412107. [Foundational paper establishing the psychological $g$-factor].
4. Gottfredson, L. S. (1997). *Mainstream Science on Intelligence: An Editorial with 52 Signatories, History, and Bibliography*. Intelligence, 24(1), 13–23. https://doi.org/10.1016/S0160-2896(97)90011-8. [Reproduces and discusses the 1994 statement signed by 52 researchers; a signatory statement, not a field consensus].
5. Russell, S., & Norvig, P. (2020). *Artificial Intelligence: A Modern Approach* (4th ed.). Pearson. ISBN 978-0134610993. [Canonical textbook framing intelligence as rational agency].
6. Chollet, F. (2019). *On the Measure of Intelligence*. arXiv preprint arXiv:1911.01547. [Critique of task-specific benchmarking and introduction of the Abstraction and Reasoning Corpus (ARC)].

7. Gardner, H. (1983). *Frames of Mind: The Theory of Multiple Intelligences*. Basic Books. ISBN 0-465-02508-0. [Source of the theory of multiple intelligences].
8. Bostrom, N. (2012). *The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents*. Minds and Machines, 22(2), 71–85. https://doi.org/10.1007/s11023-012-9281-3. [Source of the orthogonality thesis, a hypothesis].
9. Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., & Steinhardt, J. (2021). *Measuring Massive Multitask Language Understanding*. arXiv:2009.03300 (first posted 2020). https://arxiv.org/abs/2009.03300. [Defines the MMLU benchmark].
10. Jimenez, C. E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., & Narasimhan, K. (2024). *SWE-bench: Can Language Models Resolve Real-World GitHub Issues?* arXiv:2310.06770 (first posted 2023). https://arxiv.org/abs/2310.06770. [Defines the SWE-bench benchmark].
11. Brooks, R. A. (1991). *Intelligence without representation*. Artificial Intelligence, 47(1–3), 139–159. https://doi.org/10.1016/0004-3702(91)90053-M. [Argues for an embodied, behavior-based approach to intelligence].

## 12. Evidential Classification Matrix
| Section / Claim | Evidence Category | Supporting Source | Confidence / Status |
| :--- | :--- | :--- | :--- |
| Psychological $g$-Factor Existence | Well Supported (Empirical Psychology) [VERIFICATION NEEDED: no modern empirical review is cited] | Spearman (1904), Gottfredson (1997) | Moderate |
| Rational Agent Mathematical Framework | Established | Russell & Norvig (2020) | High |
| Legg–Hutter universal intelligence measure | Proposed formal definition [Hypothetical] | Legg & Hutter (2007) | Influential, contested |
| Necessity of Embodiment for General Intelligence | Contested | Brooks (1991) argues for embodiment; opposing sources [VERIFICATION NEEDED: not cited here] | Active Debate |

## 13. Revision History
- **v1.0.0** (2026-10-15): Initial authoritative reference article drafted; citations not independently verified.
- **v0.2.0** (2026-10-15): Corrected following independent fact-check audit (docs/audits/FOUND-002_audit.md); 7 findings applied, 5 items left marked for verification.