Artificial general intelligence is increasingly used to describe highly capable artificial systems. But general intelligence is not simply intelligence applied to a large number of tasks.
Artificial general intelligence has become one of the most important and least consistently defined terms in technology. It is used to describe future machines with broadly human-level intellectual abilities, highly autonomous systems capable of performing economically valuable work, and, increasingly, frontier models whose breadth of capability has become sufficient for some observers to declare that AGI has already arrived[2].
These definitions are not equivalent.
The central question is not how impressive an artificial system appears, nor how many benchmarks it can pass. It is what makes intelligence general.
Eliezer Yudkowsky offered a useful way of approaching that question long before the arrival of contemporary large language models. He described intelligence in terms of “efficient cross-domain optimisation”[1], drawing an important distinction between systems that are extraordinarily capable within a particular domain and intelligence that can operate effectively across many different domains.
A chess engine may possess capabilities vastly beyond those of any human chess player while remaining almost completely useless outside chess. Human beings are very different. We can encounter activities, technologies and intellectual disciplines for which evolution never specifically prepared us, learn how they work and apply broadly reusable cognitive abilities to them.
Humans were not born with dedicated mental machinery for computer programming, constitutional law, structural engineering or quantum mechanics. We developed those disciplines because human intelligence is sufficiently general to acquire competence in domains that did not exist when our cognitive architecture evolved.
That distinction provides a useful starting point for understanding what artificial general intelligence should mean.
General intelligence is not a catalogue of skills
Modern artificial intelligence systems possess an extraordinary range of capabilities. A single model may be able to write software, analyse financial information, interpret images, conduct research, solve mathematical problems, operate computers and communicate across many languages.
It is tempting to equate that breadth with general intelligence.
But there is an important difference between possessing a very large number of capabilities and possessing a general capacity to acquire capabilities that are missing.
An artificial system could, in principle, be trained across an enormous range of human knowledge and activity. Its repertoire might become so broad that there are relatively few common tasks on which it appears incompetent. That would make the system extraordinarily useful, but the breadth of its training would not necessarily tell us what happens when it encounters something genuinely outside that repertoire.
Generality becomes most visible at the boundary of existing competence.
Can the system encounter an unfamiliar domain, determine what matters, acquire the necessary understanding and apply that understanding effectively without its developers having to construct a new specialised system for the task?
That question gets closer to the essence of AGI.
For Eudira, a useful public definition is therefore:
Artificial general intelligence is the capacity of an artificial system to learn and apply intelligence across unfamiliar domains, rather than being confined to capabilities supplied through prior training or domain-specific programming.
The definition does not require an artificial intelligence to know everything. Humans plainly do not. Nor does it require an AGI to outperform the best human specialist in every field.
Generality is not omniscience.
It is the capacity to move beyond what is already known.
Learning is central to general intelligence
A system capable of entering unfamiliar domains must also be capable of learning.
This sounds obvious, but it becomes important when distinguishing between an artificial system that possesses a large body of pre-existing competence and one that can develop competence through experience.
Human intelligence is not static. We encounter new information, revise our understanding, develop skills and carry useful lessons from one situation into another. The person who emerges from an experience is not cognitively identical to the person who entered it.
An artificial general intelligence should possess an equivalent capacity in functional terms.
That does not mean an artificial system must learn in exactly the same way as a human brain, nor does it mean that learning must occur by changing the underlying parameters of a neural network. The implementation could take many forms.
The relevant question is simpler: does experience make the system meaningfully better equipped for what comes next?
If an artificial system encounters an unfamiliar problem, works out how to solve it and then later encounters a different problem involving a related principle, its earlier experience should be capable of improving its subsequent judgement.
Without that ability, each apparently intelligent episode remains more isolated than the behaviour of a genuinely developing intelligence.
Learning must transfer
The importance of learning becomes clearer when we consider transfer.
Suppose an artificial system solves a difficult unfamiliar problem. It explores the situation, discovers relevant relationships and eventually succeeds. That demonstrates intelligence within the task.
Now consider what happens later.
If the system encounters a different problem with a similar underlying structure, can it recognise the relationship? Can something learned previously improve the way it approaches the new situation?
This is where general intelligence differs from simple repetition.
The new problem should not need to look identical to the old one. Indeed, some of the strongest evidence of general intelligence is the ability to recognise that two superficially different situations share an underlying principle.
Humans do this constantly. A lesson learned in one workplace may change how someone understands a different organisation years later. A mathematical concept learned in one context may provide the key to solving a problem in another. Experience becomes useful beyond the circumstances in which it was acquired.
General intelligence should therefore involve more than temporary adaptation. Learning should be capable of travelling.
At the same time, intelligent transfer requires judgement. A lesson that proved useful once should not automatically be applied everywhere. The ability to recognise when previous experience is relevant is as important as the ability to retain it.
This is one reason why memory alone is not sufficient.
A system may preserve enormous quantities of information about its past without using that information intelligently. It could retrieve the right experience but draw the wrong conclusion, or apply a useful lesson in circumstances where it no longer belongs.
The existence of memory is therefore not evidence of general intelligence by itself.
What matters is what the system can do with what it remembers.
General intelligence is broader than agency
The growing ability of artificial systems to act independently has also complicated the AGI discussion.
Frontier systems can increasingly use tools, operate computers, conduct research, create software and pursue multistep objectives with limited intervention. These are substantial advances, and agency is likely to be an important characteristic of a generally intelligent system.
But autonomy and general intelligence are not interchangeable.
A specialised system can be highly autonomous while remaining narrow. Industrial control systems, automated trading systems and game-playing agents can pursue objectives effectively without possessing anything resembling general intelligence.
Agency tells us whether a system can act.
Generality tells us whether intelligence can be applied and developed across unfamiliar domains.
The two may eventually exist together in the same system, but one does not prove the other.
AGI is not the same as superintelligence
Artificial general intelligence is also frequently confused with artificial superintelligence.
They describe different properties.
Generality concerns the range and adaptability of intelligence. Superintelligence concerns its level.
A generally intelligent artificial system could, in principle, remain weaker than humans in many areas while still possessing a genuinely general capacity to learn and operate across domains. Conversely, a highly specialised artificial system might exceed every human who has ever lived within one field while remaining narrow.
This distinction matters because the arrival of AGI need not mean the immediate arrival of a machine superior to humanity in every intellectual activity.
The concepts are related, but they should not be collapsed into one another.
AGI does not require consciousness
General intelligence should also be separated from consciousness.
Whether an artificial system can possess subjective experience is an important philosophical and scientific question, but it is not necessary to resolve that question in order to assess general intelligence.
An artificial system could, in principle, learn unfamiliar domains, form useful models, solve problems, retain knowledge and transfer learning without us knowing whether there is any subjective experience associated with those processes.
Consciousness may eventually matter enormously to questions of ethics, rights and moral status. But those questions should not be silently inserted into the definition of AGI.
A capability-based definition allows the question of intelligence to be examined independently.
How would we know when AGI exists?
Ultimately, AGI should be demonstrated rather than announced.
It would not be enough for an organisation to describe a system as general, nor would the presence of components labelled memory, reasoning, planning or learning establish that general intelligence had emerged.
The evidence would need to come from behaviour.
A convincing system would demonstrate competence across substantially different domains while also showing that it could acquire abilities when its existing knowledge was insufficient. It would cope with unfamiliar problems without requiring a new specialist architecture to be constructed each time, and useful learning from one experience would be capable of improving performance elsewhere.
The tests would need to move beyond familiar benchmark questions and examine what happens when the system encounters novelty, incomplete information and circumstances that were not individually anticipated by its developers.
Human instruction should not automatically disqualify a system. Humans learn from teachers constantly. Nor should tools or external memory be treated as evidence against intelligence. Humans extend their cognition through books, computers, notes and other people.
The relevant question is not whether everything happens inside a single model.
It is whether the artificial system as a whole possesses the general capacity being claimed.
Damesh A.C. Bhiindi, founder of Eudira, argues that the same standard should apply regardless of who produces the system:
The standard cannot change depending on who built the system. Eudira should be held to it just as rigorously as OpenAI or anyone else. Describing an architecture, naming its components or announcing an ambition is not proof of intelligence. The system has to demonstrate the capability.
That principle matters because AGI is becoming commercially and culturally significant long before there is universal agreement over where its threshold lies.
If the definition becomes whatever allows the latest frontier model to qualify, the term will eventually cease to mean very much.
So, what is artificial general intelligence?
Yudkowsky’s conception of intelligence as efficient cross-domain optimisation[1] captures an essential feature of generality. Intelligence becomes general when it is no longer confined to the specific environments, problems or strategies for which it was individually prepared.
For artificial intelligence, that means more than possessing a long list of impressive abilities.
A general artificial intelligence should be capable of encountering domains outside its existing competence, learning what it needs to know and applying that learning effectively. Experience should matter beyond the immediate task, and useful learning should be capable of influencing how the system approaches unfamiliar situations later.
The distinction is simple, even if achieving it is not.
General intelligence is not the number of things an artificial system can already do. It is its ability to become competent at things it cannot yet do.