What Would Count as Artificial General Intelligence?

Artificial intelligence can already outperform humans in certain intellectual tasks. Yet exceptional ability in one area does not necessarily amount to general intelligence. What would an artificial system need to demonstrate before we could reasonably describe it as having achieved artificial general intelligence?

Artificial intelligence is advancing at a remarkable pace.

Modern AI systems can write computer programmes, analyse scientific information, solve mathematical problems and engage in sophisticated conversations. Some can perform tasks that once required years of specialist human training.

These achievements raise an increasingly important question: how would we recognise artificial general intelligence if it emerged?

The answer is less straightforward than it might appear.

What Makes Intelligence General?

Consider a world-class chess player.

They may possess an extraordinary ability to recognise patterns, anticipate consequences and develop complex strategies. Yet their mastery of chess does not necessarily mean they can repair a car, understand a medical diagnosis or learn a new language with equal proficiency.

Their intelligence, however impressive, has been developed within a particular domain.

Now imagine an artificial system that could perform all these tasks.

More importantly, imagine that it could encounter an entirely unfamiliar problem, identify what it needed to learn, acquire the necessary knowledge and apply that knowledge successfully.

This capacity to adapt across different domains brings us closer to what researchers generally mean by artificial general intelligence, or AGI.

The distinction is not simply how much an AI system knows, but how flexibly it can use what it knows.

Intelligence Beyond Individual Tasks

Most AI systems have historically been developed to perform particular categories of tasks.

A chess programme may defeat the world’s strongest players without possessing any understanding of medicine. A medical imaging system may identify abnormalities with impressive accuracy while remaining incapable of conducting an ordinary conversation.

Modern general-purpose AI systems complicate this distinction.

Large language models can perform a remarkably broad range of intellectual activities using a single underlying system. Their capabilities extend well beyond the narrow applications traditionally associated with artificial intelligence.

Nevertheless, broad competence alone does not settle whether AGI has been achieved.

An AI system might perform well across thousands of familiar tests yet struggle when confronted with unfamiliar circumstances requiring sustained adaptation, reliable reasoning or the integration of new knowledge.

The ability to succeed at many tasks is not necessarily the same as possessing general intelligence.

How Would We Recognise AGI?

There is currently no universally accepted scientific definition or definitive test for artificial general intelligence.

Some proposed definitions emphasise performance across a wide range of cognitive tasks. Others focus on learning efficiency, adaptability, autonomy or the ability to perform economically valuable work.

A convincing assessment would therefore need to examine more than impressive demonstrations.

Could the system learn unfamiliar skills without extensive retraining? Could it transfer knowledge between unrelated fields? Could it recognise its limitations, respond appropriately to uncertainty and recover from mistakes?

Could it continue to perform reliably when circumstances differed substantially from those encountered during training?

These questions are particularly important because benchmark performance does not always translate into dependable competence in the real world.

Does General Intelligence Require Consciousness?

Another distinction deserves careful consideration.

An artificial system might demonstrate extraordinary intellectual capabilities without necessarily possessing subjective experience.

Intelligence concerns capacities such as learning, reasoning, adaptation and problem-solving. Consciousness, in the sense relevant here, concerns whether there is something it is like to be that system.

Whether these properties are necessarily connected remains unresolved.

We should therefore resist assuming that an artificial system capable of general intelligence must also be conscious.

Equally, we should not dismiss the possibility without investigation.

Beyond Human Intelligence

Artificial general intelligence is sometimes discussed alongside artificial superintelligence, or ASI.

The distinction matters.

AGI generally refers to broadly capable artificial intelligence, often assessed against human cognitive abilities. ASI refers to a hypothetical form of intelligence that substantially exceeds human capabilities across a wide range of domains.

Achieving AGI would not, by definition, establish that ASI had been achieved or that its emergence was inevitable.

These are separate questions requiring separate evidence.

What Would Constitute Convincing Evidence?

Perhaps the strongest evidence of artificial general intelligence would not be a spectacular demonstration of a single capability.

It would be the consistent ability to learn, reason and adapt across unfamiliar intellectual challenges, with a degree of flexibility and reliability comparable to broadly capable human intelligence.

Even then, researchers would need to establish what had actually been measured and whether the results could be independently reproduced.

The question is not merely whether artificial intelligence can do more of what humans do, but whether it can acquire and apply knowledge with the generality, adaptability and independence that characterise human intelligence.

Until we establish clearer definitions and more rigorous methods of assessment, claims that AGI has arrived will remain open to interpretation.

And perhaps that is where the investigation must begin: not by predicting when artificial general intelligence will emerge, but by determining what would justify saying that it has.

Michael Cloke | Independent Research · Critical Inquiry · Perspectives

AI, Consciousness & Human Futures

Further Reading