Definition
Artificial general intelligence (AGI) is a general-purpose AI that can understand, learn, and carry out any intellectual task a human can, unlike today's AI, which handles only specific jobs.
A Go genius vs. a person who can do a bit of everything
The GPS in your car is amazing at finding routes, but it has no idea how to boil ramen. A Go program that beat the world champion can't place a single stone on a chess board. Programs that are good at just one thing are called "narrow AI," or specialized AI.
People are different. If you get stuck while learning to drive, you check a map. In your spare time, you pick up a new recipe. You adapt to whatever comes along. Artificial general intelligence aims to work like this: not locked into a fixed set of functions, but an intelligence that teaches itself new fields and solves problems in them.
The goal is human-level or better performance in every area human minds can reach, from writing software to giving legal advice, writing creatively, and designing science experiments.
The key difference between a smart chatbot and real intelligence
Try one of today's conversational AIs and it can feel like talking to a real person. It summarizes huge documents in seconds and even writes believable poems. So you might wonder, "Isn't AGI already here?"
But today's AI works by learning from an enormous amount of text and guessing the most likely next word. Researchers disagree about whether that counts as understanding the meaning of a sentence, and about how much it grasps of how the world works.
The common view is that it still has many limits when it comes to making a long-term plan on its own and sticking to it, or handling a surprise that wasn't in its training data as flexibly as a person would. Looking like a person and actually having the power to think are very different things.
To be more precise: the homework left to do
To be more precise, even the scientific community hasn't agreed on one absolute test for whether AGI has been reached. In the past, the "Turing test," which asks whether you can tell a machine from a person by conversation alone, was used as the yardstick. But today's language models have improved so much that they're hard to tell apart from people in short chats, so many think that yardstick alone is no longer enough.
True general intelligence would need multimodal ability, which means combining sight, hearing, and touch to understand the world. It would also need agent technology, which means setting its own goals and acting on them. Above all, AI alignment is seen as a core challenge: keeping machines from straying outside human values and safety rules.
Some also expect that once intelligence reaches human level and starts improving itself into something even smarter, it could grow explosively. That's why some people see AGI as the biggest change in human history, though views differ widely on when it might happen and what its effects would be.
Common misconceptions
The latest large language model (LLM) chatbots have already reached AGI.
Today's chatbots learn from huge amounts of data and do many things well. But many assessments say they fall short of people at planning over the long term and solving problems in a general way in unfamiliar situations. Because there's no agreed standard for AGI, though, some claim it has been reached and others say it's still far off.
Where you meet it
Even when a new programming language or accounting rule appears, it reads the manual on its own, masters it without human help, and handles the real work.
It reads medical papers, designs the molecular structure of a new drug, then directly controls lab robots and oversees clinical trials.
In one sentence
Artificial general intelligence is a general-purpose AI that isn't locked into one field and can teach itself to handle any intellectual task, the way a person can.