Defining intelligence

Most people struggle to define intelligence, and experts don’t agree either - there are over 70 definitions in scientific literature. Intelligence is often linked to reasoning, learning, and adapting, but it’s more complex than that. From IQ tests to artificial intelligence (AI), how we define and measure intelligence impacts everything from education to technology.

Here’s the challenge: traditional measures like IQ tests focus on narrow skills, while broader concepts like emotional and social intelligence remain harder to quantify. Meanwhile, AI researchers are rethinking machine intelligence, shifting from task-based skills to broader abilities like learning and reasoning in new environments.

Key insights:

• Human Intelligence: Includes reasoning, memory, creativity, emotional and social skills. Models like the Cattell-Horn-Carroll Theory break it into fluid (problem-solving) and crystallised (knowledge-based) intelligence.
• Machine Intelligence: Defined as the ability to achieve goals across different environments. New tests like ARC and ADeLe evaluate how well AI systems learn and generalise beyond training data.
• Hybrid Intelligence: Combines human judgement with AI’s processing power, creating systems where both work together to solve problems.

Thethousandproject.org offers a new perspective by focusing on intelligence as a process, not just an outcome. Their Thousand Brains Project views intelligence as emerging from interconnected learning modules, inspired by how the brain works.

This evolving understanding of intelligence highlights the need for better definitions and measurements, especially as AI continues to blur the lines between human and machine capabilities.

How Human Intelligence Has Been Studied and Measured

The Psychometric Approach: IQ, g, and Beyond

Back in 1904, Charles Spearman noticed something fascinating: people who excelled in one cognitive task often did well in others too. From this, he introduced the idea of a single underlying factor called g (general intelligence). He described it as:

“All branches of intellectual activity have in common one fundamental function (or group of functions).” [6]

This sparked debates. Thurstone countered that intelligence consists of separate abilities, a perspective that evolved into the Cattell-Horn-Carroll (CHC) Theory. This model breaks intelligence into layers, from narrow skills to broader categories. Two key components of CHC theory are fluid intelligence (Gf) - the ability to tackle new problems - and crystallised intelligence (Gc), which is based on accumulated knowledge and experience. This distinction explains why someone might shine in familiar scenarios but struggle with unfamiliar challenges, or the other way around.

As these ideas developed, neuroscience began examining how the brain’s structure and connections underpin these cognitive processes.

What Brain Research Tells Us About Intelligence

Modern neuroscience has shifted the focus from specific brain regions to how the entire brain works together. The Network Neuroscience Theory (NNT) suggests that general intelligence arises from the brain’s overall architecture and its ability to process information efficiently and flexibly:

“General intelligence reflects individual differences in system-wide mechanisms for efficient and flexible information processing.” [7]

This theory highlights the importance of long-range connections in the brain, which allow distant regions to communicate and adapt to new challenges. Two networks play key roles here: the Frontoparietal Network (FPN), which handles external problem-solving and fluid reasoning, and the Default Mode Network (DMN), which supports self-reflection and internal thought processes. Interestingly, the DMN is notably larger in humans compared to other species. Studies show that g accounts for about 58.72% of the variance in cognitive test performance [7], and people with higher intelligence often display more distinct brain network configurations when performing different tasks [8].

While cognitive abilities are central to intelligence, other dimensions like emotional and social intelligence are now being recognised as equally important.

Emotional and Social Intelligence: Expanding the Definition

Emotional Intelligence (EQ) refers to skills like empathy and self-awareness, while Social Intelligence (SI) focuses on navigating relationships effectively. According to the Social Brain Hypothesis, the human neocortex evolved not just for abstract thinking but to manage the complexities of social interactions.

A newer concept, Relational Intelligence (RQ), extends this idea further by emphasising the ability to build trust, resolve conflicts, and create shared understanding [9]. For instance, research shows that workplace relationships can influence over 30% of employee performance [9]. At Microsoft, managers who demonstrated greater empathy saw improvements in teamwork and measurable performance gains [9].

However, measuring EQ and SI is no easy task. Unlike IQ, which has well-established testing methods, assessing these forms of intelligence is more subjective. There’s ongoing debate about whether social intelligence is a separate construct or simply an application of g in social settings [10]. These broader perspectives on intelligence are particularly relevant as AI systems begin to mimic human emotional responses, complementing the technical aspects discussed in later sections.

sbb-itb-e60c46a

Every Definition of Intelligence Is Wrong. Here’s Why - Michael Bennett

How AI Researchers Define and Measure Machine Intelligence

::: @figure {Human vs Machine Intelligence: Key Differences Explained} :::

What AI Actually Is: A Working Definition

Pinning down a clear definition of artificial intelligence has always been tricky. Marvin Minsky, a pioneer in the field, once described it as:

“AI is the science of endowing machines with capabilities that, if performed by humans, would be considered intelligent.” [11]

For much of AI’s history, this meant creating systems capable of tasks like playing chess, recognising images, or translating languages. However, excelling at a single task doesn’t necessarily mean a system is intelligent. As François Chollet from Google explains:

“Solely measuring skill at any given task falls short of measuring intelligence, because skill is heavily modulated by prior knowledge and experience.” [11]

Today, many researchers turn to a broader definition proposed by Shane Legg and Marcus Hutter: “Intelligence measures an agent’s ability to achieve goals in a wide range of environments.” [2] This moves the focus from what a machine knows to how well it adjusts to different situations. It’s a perspective that has shaped how intelligence is currently measured in AI systems.

How Intelligence Is Measured in AI Systems

For years, AI performance was gauged through benchmarks like accuracy percentages, Elo ratings in games, or scores on standardised tests. These metrics mostly reflect crystallised skills but fail to capture a system’s ability to learn new abilities. Unlike traditional IQ tests, which focus on established knowledge, researchers are now shifting towards evaluating fluid intelligence - the capacity to learn and adapt quickly from minimal data.

A key example of this new approach is the Abstraction and Reasoning Corpus (ARC), introduced by François Chollet in 2025. ARC challenges systems with novel visual puzzles that demand true generalisation instead of relying on patterns memorised during training. [11]

In April 2026, a study in Nature explored this idea further using the ADeLe battery v1.0, which includes 16,108 tasks across 20 benchmarks. This study assessed 15 large language models (LLMs) and revealed some fascinating insights. While larger models excelled in knowledge-based tasks, inference-heavy systems like DeepSeek-R1 and OpenAI o1 outperformed others in logical reasoning and quantitative challenges. [12] This highlights that a model’s reasoning process is just as important as the volume of data it has been trained on.

Human Intelligence vs Machine Intelligence: Key Differences

One of the biggest distinctions between human and machine intelligence lies in generalisation. Humans excel at extreme generalisation - the ability to adapt to entirely new domains with just a few examples. In contrast, most AI systems are limited to local generalisation, meaning they can handle new data within a specific task but struggle significantly when faced with unfamiliar scenarios. [11] This underscores why evaluating fluid intelligence is crucial for both humans and AI.

Humans also benefit from evolutionary and developmental priors - innate assumptions about the world that enable rapid learning. On the other hand, AI systems often require vast amounts of data to achieve what a human can learn in just a few experiences. [11] To better understand these differences, here’s a breakdown of the three levels of generalisation:

Generalisation Level	What It Means	Example
Local (Robustness)	Can handle new data within a familiar task	ImageNet image classifiers [11]
Broad (Flexibility)	Can adapt across related tasks	Level 5 self-driving vehicles [11]
Extreme (Generality)	Can tackle entirely new domains	Human cognition [11]

This constant redefinition of intelligence, often referred to as the “AI Effect”, reflects a shifting goalpost: once a task is mastered, it is no longer seen as a true measure of intelligence. [11] Recognising these differences is essential for understanding what intelligence means in a world where human and machine capabilities increasingly overlap.

New Directions in Intelligence Research

Hybrid Intelligence: When Humans and AI Work Together

The way humans and AI collaborate is shifting. AI is no longer just a tool - it’s becoming a partner in what researchers call co-intelligence. This approach allows AI to interpret intent, consider options, and execute tasks, while humans provide guidance, values, and accountability. As Accenture Research explains:

“Use of AI is shifting from tools to co‑intelligence, where humans lead while AI amplifies judgment, execution and autonomy.” [15]

One example of this shift comes from a year-long study at the University of Southern California. Researchers worked with 10 clinicians and 105 stroke survivors, using a Dynamic Bayesian Network to assist in physical therapy assessments. The system achieved 90.6% agreement with therapists, improving decision-making in clinical settings [14]. This aligns with what Tamim Ahmed and Thanassis Rikakis describe as augmented intelligence: “the long‑term interrelated development of human and computational intelligence across environment, actions, goal setting, and meaning extraction.” [14]

Despite these advancements, adoption remains slow. Only 11% of organisations are effectively integrating systems that allow for continuous co-learning between humans and AI [15]. The challenge lies in keeping up with the rapid pace of technological development.

System-Level Intelligence in AI Platforms

Expanding on co-intelligence, researchers are delving into how intelligence emerges from interconnected systems. A key idea from thethousandproject.org suggests that intelligence doesn’t reside in a single algorithm but arises from how different components interact. This concept is central to the Thousand Brains Project’s open-source framework, Monty. Inspired by the human neocortex, Monty uses “learning modules” based on over 1,000 cortical columns. These modules communicate through a Cortical Messaging Protocol (CMP) and reach decisions via a voting algorithm, avoiding reliance on a centralised system [19].

Monty also offers impressive efficiency. It uses 33,000 times less computation than a standard Vision Transformer for training and 527 million times less than pretraining plus fine-tuning [4]. As Viviane Clay of Numenta explains: “The power of the mammalian brain lies in its reuse of cortical columns as primary computational units.” [3]

This underscores an important principle: intelligent behaviour depends more on system architecture than on sheer computational power. Nic Windley summarises this idea: “Intelligence isn’t a switch that flips on. It’s something that emerges.” [16]

However, as these systems grow more sophisticated, figuring out how to measure their intelligence introduces new hurdles.

The Problem of Measuring New Forms of Intelligence

As intelligence evolves into more distributed and hybrid forms, traditional measurement tools fall short. Current benchmarks can assess what a system can do today but say little about its reliability in new situations. For instance, state-of-the-art AI systems score 0% on the ARC-AGI 3 test, which measures the ability to learn concepts in unfamiliar environments [18].

To address these gaps, new frameworks are being developed. The Cognitive and Artificial Intelligence Evaluation (CAIE) framework assesses over 90 cognitive features across six categories: Learning, Perception, Reasoning, Interaction, Memory, and Optimisation [13]. Another approach, the machine Perturbational Complexity and Agency Battery (mPCAB), takes a different route. Instead of focusing on task performance, it uses intentional system modifications to uncover causal links between internal mechanisms and behaviour, moving beyond surface-level correlations [17].

One persistent challenge is the “mimicry” problem: how to differentiate genuine understanding from advanced pattern recognition. Until we can reliably measure this distinction, any claims about a system’s intelligence remain uncertain at best.

Conclusion: What Intelligence Means in the Age of AI

Key Takeaways

Intelligence has always been a rich and layered concept. Over time, researchers have proposed many definitions, with a 1995 report from the American Psychological Association revealing 24 different interpretations from 24 theorists [1][5]. This lack of consensus becomes even more pronounced as AI capabilities continue to evolve.

Traditional measures like IQ only scratch the surface, especially in an age where human and machine cognition are deeply intertwined. Intelligence today is increasingly seen as something that emerges from complex, interconnected systems, rather than being confined to individuals. This perspective is a key focus of thethousandproject.org, which emphasises how system-level architecture - rather than individual processing power - drives intelligent behaviour.

The idea of an Extended Intelligence Quotient (X-IQ) reflects this shift. Human cognition, when amplified by AI tools, digital memory, and sensors, creates a blend of individual and collective thought [21]. As Sharon Gal-Or aptly puts it:

“The future will not belong only to smarter individuals, but to smarter systems.” [21]

These evolving ideas challenge us to rethink intelligence, raising profound questions about understanding and decision-making.

Open Questions Worth Asking

Some of the most pressing questions in this field remain unresolved. Can machines ever achieve true comprehension, or are they destined to remain advanced pattern-matchers? Philosopher Daniel Dennett highlights this dilemma, describing how AI often demonstrates “competence without comprehension” [20] - delivering results without truly understanding them.

On a broader scale, there’s a challenge for society as a whole. Jakub Bareš from the European Nexus for Strategic Intelligence puts it succinctly:

“When solutions become cheap, the scarce skill is deciding what deserves to be solved at all.” [22]

As AI accelerates our ability to pursue goals, the risk of poorly defined goals looms larger. This underscores the importance of sound judgement and ethical decision-making. Researchers often refer to this as “upstream leverage” - the ability to determine which problems are truly worth addressing. In this context, the most vital form of intelligence in the age of AI might still be a distinctly human quality: the capacity to decide what truly matters.

FAQs

Is IQ actually a good measure of intelligence?

IQ tests evaluate how well someone can tackle certain abstract, verbal, and spatial reasoning problems. They’re often used to gauge potential for academic and professional achievements. However, they fall short when it comes to capturing qualities like creativity, emotional understanding, or the ability to make thoughtful decisions. While IQ tests are helpful for measuring specific cognitive abilities, they don’t encompass the broader, multifaceted nature of human intelligence.

Can AI ever truly understand, not just mimic?

AI’s ability to truly understand or merely mimic human cognition is a hotly debated topic. Modern AI excels at tasks like pattern recognition and logical reasoning, but it doesn’t possess the direct experience or ongoing thought processes that are central to human understanding. While some suggest that AI demonstrates a kind of generalisation, others argue it simply follows pre-learned rules. Unlike humans, AI operates reactively and within isolated sessions, lacking the continuous, self-driven thinking that defines the human mind.

What is hybrid intelligence in practice?

Hybrid intelligence brings together the strengths of human and artificial intelligence, creating a partnership that achieves outcomes neither could reach independently. Instead of replacing human judgement, these systems amplify it by combining the speed and efficiency of AI with the intuition, ethical reasoning, and decision-making skills unique to humans.

In practical terms, this collaboration works through iterative feedback loops. AI takes the lead in analysing vast amounts of data, while humans step in to ensure accountability and provide critical oversight. Workflows are designed so machines handle repetitive, large-scale tasks, leaving humans to focus on adding context, making nuanced decisions, and steering the process in the right direction. Together, this dynamic duo delivers results that are both efficient and thoughtful.