AGI, short for artificial general intelligence, means an AI system that can handle a wide range of thinking tasks at roughly human level or better, instead of excelling at one narrow job. There is no single agreed definition: OpenAI, Google DeepMind, and Anthropic each describe it differently, and researchers disagree about whether today's methods will get there.
That disagreement is the most important thing to understand about AGI. When someone says AGI is close, far away, or already here, the first question to ask is which definition they are using. This guide lays out the main ones, where they came from, and what they mean for anyone using AI today.
What Does AGI Mean?
The key word is general. Most AI in real use is what researchers call narrow AI: a system built for a specific task, such as recognizing speech, recommending products, or translating text. It can be excellent at that task and useless outside it. Our glossary entry on AGI puts it simply: a hypothetical AI that can learn and perform any intellectual task a human can, rather than being good at one narrow thing.
General intelligence, in the way people use the phrase, has three parts:
- • Breadth. It works across many kinds of tasks, not just one.
- • Depth. It performs those tasks well, at least as well as a capable person.
- • Learning. It can pick up new skills it was not specifically built or trained for.
Today's large language models blur the line, because one model can write, summarize, code, and answer questions on almost any topic. That breadth is why the AGI debate got louder after tools like ChatGPT arrived. Whether breadth alone counts as general intelligence is exactly where the definitions split.
Where Did the Term AGI Come From?
According to TIME's profile of Google DeepMind co-founder Shane Legg, the term AGI was first used by physicist Mark Gubrud in 1997. In 2002, Legg independently came up with the phrase and popularized it with his former boss, Ben Goertzel.
The label is newer than the ambition behind it. What the term added was a clear way to separate the goal of broad, human-like intelligence from narrow AI built for one task.
How Do the Major AI Labs Define AGI?
The companies building the most advanced models do not agree on a definition, and the differences are revealing.
- • OpenAI defines AGI in its Charter as "highly autonomous systems that outperform humans at most economically valuable work." That definition is about economic capability: what the system can do, measured against the work people do.
- • Google DeepMind researchers proposed a framework instead of a single line. Their paper, Levels of AGI for Operationalizing Progress on the Path to AGI by Meredith Ringel Morris, Shane Legg, and colleagues, published in the Proceedings of ICML 2024, rates systems on two axes: performance (how well) and generality (how broadly).
- • Anthropic CEO Dario Amodei avoids the term. In his October 2024 essay Machines of Loving Grace, he writes that he finds AGI "an imprecise term that has gathered a lot of sci-fi baggage and hype," and uses "powerful AI" instead. He describes powerful AI as smarter than a Nobel Prize winner across most relevant fields, and summarizes the idea as "a country of geniuses in a datacenter."
These are not small wording differences. OpenAI's definition centers on outperforming humans at most economically valuable work. DeepMind's is a ladder with several steps. Amodei's is a specific description of capability that skips the label entirely.
What Are the Levels of AGI?
The DeepMind framework is useful because it replaces a yes-or-no question with a scale. For general AI, the paper defines five performance levels above "no AI":
- • Level 1, Emerging: equal to or somewhat better than an unskilled human.
- • Level 2, Competent: at least the 50th percentile of skilled adults.
- • Level 3, Expert: at least the 90th percentile of skilled adults.
- • Level 4, Virtuoso: at least the 99th percentile of skilled adults.
- • Level 5, Superhuman: outperforms 100 percent of humans.
In the paper's own table, systems including ChatGPT, Bard, Llama 2, and Gemini were placed at Level 1, Emerging AGI. The authors also map other definitions onto their scale: they write that OpenAI's labor-replacement threshold better matches Virtuoso AGI. In other words, by one framework today's chatbots are an early form of AGI, and by another the bar sits several levels higher. Both statements can be true because they use different definitions.
The paper also sets out six principles, including focusing on what a system can do rather than how it works inside, and treating progress as a path with stages rather than a single finish line.
How Do Researchers Define General Intelligence?
Outside the labs, researchers have tried to define intelligence itself, which turns out to be the harder problem.
In a 2007 paper in Minds and Machines, Universal Intelligence: A Definition of Machine Intelligence, Shane Legg and Marcus Hutter gathered many expert definitions of intelligence and distilled them into one informal working definition: "Intelligence measures an agent's ability to achieve goals in a wide range of environments." The phrase "wide range of environments" is what makes it general.
In 2019, AI researcher François Chollet argued in On the Measure of Intelligence that measuring skill at specific tasks falls short, because enough prior knowledge or training data can buy high skill without real generalization. He defined intelligence as skill-acquisition efficiency: how efficiently a system learns new skills, not how many it already has. He also introduced the Abstraction and Reasoning Corpus (ARC), a benchmark designed to test that kind of fluid, human-like generalization.
Under Chollet's view, a system that has memorized enormous amounts of knowledge is not necessarily general. Under a capability-focused view like DeepMind's, what matters is what the system can do. That tension runs through every debate about whether a new model is or is not AGI.
Is Today's AI Already AGI?
It depends on the definition, and the research community is far from convinced. In March 2025, the Association for the Advancement of Artificial Intelligence published its AAAI 2025 Presidential Panel report on the Future of AI Research, which included a survey of 475 respondents. The majority, 76 percent, said that "scaling up current AI approaches" to yield AGI is "unlikely" or "very unlikely" to succeed.
That does not mean progress has stalled. It means many researchers doubt that simply making today's models bigger will produce general intelligence on its own, and expect new ideas will be needed. Meanwhile, the same report found that 70 percent of respondents opposed halting AGI research until full safety and control mechanisms are in place.
A fair summary: current models are remarkably broad, uneven in depth, and still make mistakes no capable person would make. Some definitions call that early AGI. Most researchers in the AAAI survey do not see a straight line from here to the full version.
How Is AGI Different From an AI Agent?
These two terms get mixed up constantly. An AI agent is software that pursues a goal and takes actions to reach it, using tools, memory, and step-by-step reasoning. Agents exist today, and businesses use them to answer calls, book appointments, and handle routine work. Our plain-English guide to AI agents explains how they work.
AGI describes a level of capability. An agent describes a way of using AI. A narrow agent can be very useful without being anywhere close to AGI, and a hypothetical AGI would not have to be built as an agent at all. When a product calls itself an "AGI agent," it is worth asking what, specifically, it can do.
When Will AGI Arrive?
Nobody knows, and anyone who states a date as fact is guessing. Predictions from researchers and company leaders range widely, and they depend on which definition is being used. Even people who expect fast progress hedge heavily. In Machines of Loving Grace, Amodei writes that his predictions are "unavoidably going to consist of guesses" and that "everything I'm saying could very easily be wrong."
The more useful question for most people is not when, but what changes as AI gets more capable, step by step. The DeepMind levels framework was built for exactly that: tracking progress in stages instead of waiting for a single moment.
What Does AGI Mean for a Small Business Today?
For day-to-day decisions, AGI is mostly a distraction. The tools available right now are narrow or broad-but-uneven systems, and they deliver value when they are pointed at specific, well-defined jobs with a human checking the output.
- • Ignore AGI marketing. A product's usefulness depends on what it does for your workflow, not on whether it claims to be general.
- • Judge tools by tasks. Test any AI tool on the actual jobs you need done, and check the results.
- • Keep humans in charge of judgment calls. Even the most capable current models get things confidently wrong, so anything involving money, customers, or safety still needs review.
- • Stay curious, not anxious. Following how the definitions and evidence evolve is worthwhile. Planning your business around a date nobody can predict is not.
Sources
- • TIME, Shane Legg (TIME100 AI)
- • OpenAI Charter
- • Morris et al., Levels of AGI for Operationalizing Progress on the Path to AGI, ICML 2024 (arXiv)
- • Dario Amodei, Machines of Loving Grace (October 2024)
- • Legg and Hutter, Universal Intelligence: A Definition of Machine Intelligence, Minds and Machines, 2007 (arXiv)
- • Chollet, On the Measure of Intelligence, 2019 (arXiv)
- • AAAI 2025 Presidential Panel on the Future of AI Research (PDF)
Tell us what is costing you the most time. We will map out exactly what your business needs. Free, no obligation.
Build An Agent