What is artificial general intelligence?
Artificial general intelligence (AGI) is the name for a hoped for kind of AI that could learn and reason across almost any task a person can, rather than excelling at one narrow job. An AGI could pick up a new subject, apply what it knows in unfamiliar situations and solve problems it was never specifically built for.
Today's AI, including the models behind AI Chat and our AI agents, is often called narrow AI. It can be remarkably good at writing, coding, translating or spotting patterns in images, yet it still has gaps that people do not: it can fail at simple common sense, struggle to learn from a handful of examples and lose track over long, open ended tasks.
AGI is sometimes called strong AI, full AI or human level AI. There is no agreed definition and no agreed test, so claims that a system "is AGI" should always be treated with care.
How AGI differs from AI and superintelligence
It helps to place AGI on a simple ladder:
| Term | What it means |
|---|---|
| Narrow AI | Systems built for specific tasks. This is all AI in use today. |
| Artificial general intelligence | A system as broadly capable as a person across most kinds of thinking. Hypothetical. |
| Artificial superintelligence | A system that far exceeds the best human minds in nearly every field. Hypothetical. |
AGI is often seen as a possible step towards superintelligence, but nothing guarantees one leads to the other. Our Superintelligence Wiki covers that next step in depth, and the AI Wiki explains how today's narrow AI works.
What would count as general intelligence?
Researchers list a number of abilities a truly general system would need:
- Reasoning and solving problems, including under uncertainty.
- Knowledge of the world, including the everyday common sense people take for granted.
- Planning towards goals over many steps.
- Learning new skills quickly from little data, and keeping old skills while doing so.
- Communicating in natural language.
- Combining all of these to reach a goal.
Many people add the ability to perceive the world through sight and sound, and to act in it, for example by controlling a robot or using a computer. Some also argue for traits such as imagination and the ability to explain its own thinking.
How would we know AGI has arrived?
Because there is no single definition, many tests have been proposed. Each one captures part of the picture:
- The conversation test. Proposed by Alan Turing in 1950, a machine passes if a judge cannot reliably tell its written replies from a person's. Modern chat models can often pass short versions, which shows the test says more about convincing conversation than about general intelligence.
- The exam test. A system sits university level exams across many subjects and passes them. Large language models already score highly on many such exams.
- The coffee test. A robot walks into an unfamiliar home and makes a cup of coffee, finding the kitchen, the machine, the cups and the water by itself.
- The employment test. A system does an ordinary job as well as a person, for a sustained period, across a range of roles.
- The flat pack test. A machine reads the instructions and assembles furniture from parts.
Some researchers instead propose levels of AGI, scoring systems on how general and how capable they are, much like the levels of autonomy used for AI agents in our AI Agent Wiki.
Hard problems on the way to AGI
Some tasks are so broad that solving them would seem to require general intelligence. They are sometimes called AI complete problems. Examples include truly understanding any piece of writing, translating with full grasp of meaning and culture, and handling the endless surprises of the physical world. Modern models have made rapid progress on many of these, which is one reason the AGI debate has become so lively.
A short history of the idea
The founders of AI in the 1950s and 1960s aimed for general intelligence from the start and expected to reach it within a generation. Early programs could prove theorems and play games, but progress slowed when researchers met the huge difficulty of common sense, perception and learning. Funding dried up more than once in periods now known as AI winters.
From the 1990s attention shifted to narrow AI that could deliver useful results, such as search, recommendations and speech recognition. A small community kept working on general intelligence, and the term artificial general intelligence became widely used in the early 2000s to separate that goal from narrow AI.
Deep learning changed the picture in the 2010s, and large language models changed it again in the 2020s. Systems that could write, code, reason through problems and use tools led some researchers to say that early signs of general intelligence were appearing, while others argued these systems still lack real understanding. Several leading AI companies now state openly that building AGI is their goal. Read how conversational AI developed in AI chat before ChatGPT.
When might AGI arrive?
Nobody knows. Surveys of AI researchers have produced a wide spread of answers, from within a few years to many decades away, and a minority believe it may never happen. Forecasts have moved much earlier since large language models appeared. Bold predictions about AI have often been wrong in both directions, so any firm date should be treated with caution. Our article on how far away superintelligence is looks at the same question one step further on.
Routes researchers are exploring
There is no agreed path to AGI. The main ideas include:
- Scaling. Making today's models bigger and training them on more data and computing power.
- New architectures. Adding memory, planning and reasoning methods that current models lack.
- Agents and tools. Letting models act, use software and learn from the results, as our Super Intelligence AI Agent does in a controlled way.
- Embodiment. Putting AI into robots so it learns from the physical world.
- Learning like the brain. Drawing on neuroscience, up to the idea of copying how a whole brain works in software, which remains far beyond current science.
Hardware matters too. Some researchers look to new kinds of computing, including quantum computing, explained in the AI Wiki.
Does AGI need consciousness?
Not necessarily. Being able to do a task is a different question from having experiences or feelings while doing it. A system could in principle match people on every test without being conscious at all, and we have no reliable way to measure consciousness in a machine. Philosophers have long debated whether a machine that acts intelligently truly understands, or only seems to. These questions matter for how we treat AI and how we judge its claims about itself.
Possible benefits
If AGI were built and used well, supporters believe it could speed up science and medicine, help solve problems such as disease and climate change, raise living standards and give everyone access to expert help in education, health and law. Even today's narrow AI already shows a small part of this, from AI tools that write and design to agents that handle routine admin.
Risks and concerns
AGI also raises some of the most serious questions in technology:
- Control and alignment. Making sure a highly capable system pursues the goals we intend, and keeps doing so as it improves. Read our research on robust AI security and alignment.
- Misuse. Powerful systems could help bad actors with cyberattacks, weapons or large scale fraud.
- Jobs and the economy. A system able to do most work could reshape employment and widen inequality unless the benefits are shared.
- Concentration of power. The few organisations able to build AGI could gain enormous influence.
- Existential risk. Some leading researchers warn that losing control of systems smarter than us could threaten humanity itself, while others see such fears as overstated or as a distraction from present day harms.
Many experts argue that AI agents are where these risks first become real, because an agent can act, not just advise. That is why every agent on our platform follows our AI governance framework and Zero Trust AI Security, and why laws such as the EU AI Act matter.
AGI and the Ask AI course
AGI comes up throughout our Certified Applied AI Professional course. Use this wiki when a question asks how AGI differs from today's AI, how it might be tested, or what risks it brings. For definitions of the terms used here, see the AI Dictionary and the AI Dictionary for Everyone.
Frequently asked questions
Does AGI exist today?
No. Today's systems are impressive and increasingly broad, but there is no agreed evidence that any of them matches people across nearly all kinds of thinking.
Is ChatGPT an AGI?
No. ChatGPT and similar models are powerful narrow AI. They can do many tasks well but still make basic mistakes, cannot reliably learn new skills on the spot and depend on people to check their work.
Is AGI the same as superintelligence?
No. AGI means roughly human level ability across the board. Superintelligence means going far beyond it. See the Superintelligence Wiki.
How is AGI different from an AI agent?
An AI agent is software that pursues a goal and uses tools. Agents today run on narrow AI. AGI describes how general the intelligence is, not whether it acts.
Selected references
- AGI-01 Alan Turing, "Computing Machinery and Intelligence", Mind, 1950.
- AGI-02 Ben Goertzel and Cassio Pennachin (editors), Artificial General Intelligence, Springer, 2007.
Related research papers
Papers in our research section that go deeper on topics covered here.