What is artificial general intelligence (AGI)?

What AGI means

Artificial general intelligence names an AI that possesses general-purpose intelligence comparable to a human's: the ability to learn a wide range of tasks, reason about unfamiliar problems, transfer knowledge from one domain to another, and adapt without being specifically retrained for each new job. IBM defines AGI as a hypothetical stage in AI development where a system "can match or surpass human cognitive capabilities across any task."

The defining word is general. Humans use one mind to cook, argue a legal case, learn a language, and fix a bicycle. An AGI would show that same breadth and flexibility rather than excelling at a single predefined function. This is what separates the concept from the impressive but bounded systems that exist now.

It is important to state plainly that AGI is a goal and a research concept, not an existing product. No system today qualifies, and the term is used to describe a destination that different researchers define, and expect, in very different ways.

AGI versus the narrow AI we have today

Every AI system in use today is narrow AI (also called weak AI): it is built and trained for a specific set of tasks, however broad those tasks may appear. Image classifiers, recommendation engines, self-driving stacks, and even large language models are, in this framing, narrow, they operate within the boundaries of what they were trained to do and do not possess general understanding or autonomy across arbitrary domains.

Modern large language models complicate the picture because they are strikingly versatile, writing code, drafting essays, answering questions across countless subjects. This breadth has led some researchers to argue they represent early steps toward general intelligence. But they still fail in ways that reveal their limits: they hallucinate facts, struggle with reliable multi-step reasoning, lack persistent memory and grounding in the physical world, and can be brittle on tasks slightly outside their training distribution.

The mainstream position, reflected by IBM, Stanford HAI, and others, is that current systems, however capable, remain narrow AI. Versatility within language and pattern tasks is not the same as the robust, reliable, cross-domain generality that AGI implies.

Why the definition is contested

There is no single, agreed-upon definition of AGI, and this is the root of much of the debate. Different framings set the bar in different places. An early, informal standard was the ability to perform any intellectual task a human being can. OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work," an economic rather than cognitive yardstick. Others tie it to passing specific tests, though classic proposals like the Turing test are now widely seen as inadequate measures of general intelligence.

To bring order to this, researchers at Google DeepMind proposed a framework in the 2023 paper "Levels of AGI: Operationalizing Progress on the Path to AGI" (Morris et al.). It grades systems along two axes, performance and generality, into levels from Emerging to Competent, Expert, Virtuoso, and Superhuman. Notably, the authors classify today's leading frontier language models as "Emerging AGI": general but only matching or modestly exceeding unskilled humans on a broad range of tasks, and far from the higher tiers.

Because the finish line itself is disputed, claims that AGI is near or far often reflect which definition the speaker is using as much as any fact about the technology.

The debate over how close we are

Views among serious researchers span an unusually wide range. Some leaders at frontier labs argue that systems with general capabilities could arrive within years and that current scaling trends point toward it. Others, including many academics, contend that today's approaches are missing fundamental ingredients, robust reasoning, causal understanding, grounding, continual learning, and that AGI, if it comes, is decades away or may require paradigms not yet invented.

A flashpoint in this debate was the 2023 Microsoft Research paper "Sparks of Artificial General Intelligence: Early Experiments with GPT-4" (Bubeck et al.), which argued that GPT-4 exhibits early signs of general intelligence. The paper drew intense criticism for relying on informal observation rather than rigorous, reproducible benchmarks, and for a definition that critics found too loose. It stands as a good illustration of why AGI claims are so contentious: impressive demonstrations are read very differently depending on one's definition and standards of evidence.

Stanford's Institute for Human-Centered AI and other academic voices tend to counsel caution, emphasizing that current systems, despite rapid progress, still lack the reliability and generality that AGI denotes, and that hype can outrun measured evidence.

What would AGI require, and how would we know?

Researchers commonly list capabilities an AGI would need that current systems lack or handle poorly: reasoning through genuinely novel problems, learning efficiently from little data, transferring skills across domains, maintaining and using long-term memory, understanding cause and effect, and acting reliably rather than confidently guessing. Some argue physical grounding or embodiment matters; others focus on purely cognitive benchmarks.

Measurement is itself unsolved. No consensus test exists that a system could pass to be declared an AGI, and older proposals are widely considered insufficient. This is precisely the gap the DeepMind "Levels of AGI" framework tries to fill, by measuring performance and generality against percentiles of skilled adults across many tasks rather than seeking a single pass/fail moment.

The honest summary is that we lack both an agreed definition and an agreed yardstick, which is why responsible sources describe AGI as a research goal surrounded by real uncertainty rather than an imminent, well-specified milestone.

AGI, superintelligence, and why it matters

AGI should not be confused with artificial superintelligence (ASI), a further hypothetical stage describing an intellect that vastly surpasses the best human minds across essentially all domains. AGI is roughly human-level generality; ASI is far beyond it. Both are speculative, and ASI is even more so.

The concept matters regardless of timeline because the prospect of highly general, autonomous systems drives major decisions today, in research funding, corporate strategy, and public policy, and motivates the field of AI safety and alignment, which studies how to keep increasingly capable systems reliable and controllable. Discussions of AGI risk and governance are therefore substantive even while AGI itself remains unrealized.

The balanced view to hold: AGI is a genuine, actively pursued research goal, not science fiction to be dismissed nor an accomplished fact to be assumed. Current AI is narrow, progress is real and fast, and both the definition and the arrival of AGI remain open questions on which informed experts disagree.

Frequently asked questions

Has AGI been achieved?

No. There is broad agreement among mainstream sources, including IBM and Stanford HAI, that AGI does not yet exist. Today's AI systems, including the most advanced large language models, are narrow AI: powerful within their training but lacking the reliable, cross-domain general intelligence that AGI denotes. Some researchers argue frontier models show early signs of generality, but that remains a contested interpretation, not a consensus that AGI has arrived.

What is the difference between AGI and narrow AI?

Narrow AI (or weak AI) is built for a specific set of tasks and does not generalize beyond them; all AI in use today is narrow, including image recognizers and language models. AGI is a hypothetical system with broad, human-level general intelligence that can learn and reason across virtually any domain, transferring knowledge flexibly the way a person does. The gap is generality and reliability, not raw capability on any single task.

When will AGI be achieved?

Nobody knows, and expert estimates vary enormously, from a few years to several decades to possibly never. The disagreement is genuine and stems partly from the lack of an agreed definition of AGI and partly from differing views on whether current methods can scale to general intelligence or are missing fundamental ingredients. Any confident single date should be treated with skepticism; the responsible framing is deep uncertainty.

Are large language models like GPT-4 a form of AGI?

Not in the mainstream view. LLMs are remarkably versatile, which has led some, such as the authors of Microsoft's 2023 "Sparks of AGI" paper, to argue they show early signs of general intelligence. But that claim was heavily criticized for weak evidence, and LLMs still hallucinate, reason unreliably over multiple steps, and lack memory and grounding. DeepMind's "Levels of AGI" framework classifies them at most as "Emerging AGI," far from human-level general intelligence.

What is the difference between AGI and superintelligence (ASI)?

AGI (artificial general intelligence) refers to AI with roughly human-level general intelligence across domains. Artificial superintelligence (ASI) is a further, more speculative stage: an intellect that vastly exceeds the best human minds in essentially every field. AGI would match humans broadly; ASI would dramatically surpass them. Both are hypothetical and do not exist today, with ASI being the more distant and uncertain concept.

Why is AGI so hard to define?

Because intelligence itself is hard to pin down, and different framings set the bar in different places, some cognitive (matching human reasoning across tasks), some economic (OpenAI's "outperform humans at most economically valuable work"), some test-based. There is also no agreed measurement that a system could pass to be declared an AGI. DeepMind's "Levels of AGI" framework tries to address this by grading systems on performance and generality rather than seeking a single pass/fail moment.

Related

Explainers: LLM, GenAI, Diffusion model

Glossary: Superintelligence (ASI), AI alignment, Machine learning

Sources

• IBM Think — What is artificial general intelligence (AGI)?
• Morris et al. (2023), Google DeepMind — Levels of AGI: Operationalizing Progress on the Path to AGI
• Bubeck et al. (2023), Microsoft Research — Sparks of Artificial General Intelligence: Early Experiments with GPT-4
• Stanford Institute for Human-Centered AI (HAI)
• Wikipedia — Artificial general intelligence