Edited by 
Short answer
AGI stands for artificial general intelligence: an AI system that can understand, learn and carry out most intellectual tasks a capable human can, including ones it was never specifically trained for. Today's AI is mostly narrow: excellent at certain tasks but unreliable outside them. There is no single agreed definition of AGI, so experts disagree on whether it has arrived; OpenAI defines it as "highly autonomous systems that outperform humans at most economically valuable work." AGI is seen as the step before superintelligence (ASI), AI that far surpasses humans.
What does AGI mean?
The "general" in artificial general intelligence is the key word. A chess engine is superhuman at chess and useless at anything else. A human can learn chess, cooking, law and carpentry with the same brain. AGI would be an AI with that kind of flexibility: able to transfer what it knows to new problems, learn new skills from a few examples and work towards long-term goals on its own.
The term became popular in the early 2000s, when researchers including Ben Goertzel and Shane Legg, later a co-founder of Google DeepMind, used it to separate the original goal of AI research, human-like general intelligence, from the narrow applications the field had become known for.
AGI vs today's AI
| Narrow AI (ANI) | Today's frontier models | AGI | |
|---|---|---|---|
| Range | One task or domain | Very broad: writing, coding, maths, images, speech | Almost all intellectual tasks |
| Learning new skills | Needs retraining | Some, from instructions and examples | Learns like a person, from little data |
| Reliability | High within its task | Uneven: brilliant on some tasks, basic mistakes on others | Consistently at least human level |
| Working alone on long tasks | No | Improving, with AI agents | Yes |
Large language models such as ChatGPT, Claude and Gemini blur the line. They are general in range but still make errors a person would not, struggle with long open-ended projects and do not reliably learn from experience after training. Read how AI works to see why.
Common definitions of AGI
- OpenAI: "highly autonomous systems that outperform humans at most economically valuable work." Under OpenAI's 2025 agreement with Microsoft, a claim that AGI has been reached is to be verified by an independent expert panel.
- Google DeepMind's "Levels of AGI": rates systems from Emerging to Competent, Expert, Virtuoso and Superhuman, by both how well and how generally they perform. On this scale, today's chatbots are "Emerging AGI".
- Capability view: a system that can do any cognitive task an average adult can do, including learning new ones.
- Economic view: AI that can do most remote, computer-based jobs as well as people.
Because the definitions differ, two experts can look at the same model and disagree about whether it is AGI.
How do you test for AGI?
| Test | What it checks |
|---|---|
| Turing test (1950) | Whether a machine can hold a conversation indistinguishable from a human's. Modern chatbots can often pass short versions, which is why it is no longer seen as enough. |
| Coffee test (Steve Wozniak) | Whether a robot can walk into an unfamiliar home and make a cup of coffee: a test of general physical and practical intelligence. |
| ARC-AGI (François Chollet) | Visual puzzles that are easy for people but hard for AI, testing the ability to learn new rules from a few examples. The ARC Prize offers money for systems that reach human-level scores. |
| Real-world task benchmarks | How long and how complex a task AI agents can complete on their own, such as software engineering or research tasks. |
Has AGI been achieved?
Not by most definitions, as of late 2026. Frontier models now beat most people on many exams, coding tests and science questions, and some researchers argue that by a broad definition we already have early AGI. Others point to gaps that remain: reliability, learning on the job, long-term planning and physical-world skills. AI lab leaders have predicted AGI within a few years; many academic researchers expect it to take longer. Read when will AGI arrive? for the main forecasts.
What would AGI need?
- Compute: training frontier models already uses tens of thousands of GPUs or TPUs. AGI-level systems are expected to need far more, which is why companies are building gigawatt AI data centers. Read what is an AI data center.
- Algorithms: better ways to learn from less data, to reason over many steps and to remember and improve over time.
- Data and experience: high-quality data, plus learning from interacting with the world, not only from text.
- Safety: reliable ways to make sure a general system does what we intend. This is the focus of alignment research at labs such as SSI, Anthropic, OpenAI and Google DeepMind.
AGI and superintelligence
AGI matches humans; superintelligence (ASI) would far exceed them. Many researchers think the step from one to the other could be short, because an AGI good at AI research could help build its own successor. Read AGI vs ASI, what is superintelligence and the AI 2027 scenario for how that might unfold.
Frequently asked questions
What does AGI stand for?
Artificial general intelligence. In finance and taxes, AGI also stands for adjusted gross income, which is unrelated.
Is ChatGPT an AGI?
Not by most definitions. It is a general-purpose model, but it still makes basic mistakes, cannot reliably learn new skills after training and struggles with long, open-ended tasks. Google DeepMind's framework would call it "Emerging AGI".
What is the difference between AI and AGI?
AI is the whole field of making machines do tasks that need intelligence, and almost all AI today is narrow. AGI is the goal of AI that is generally capable across nearly all intellectual tasks, like a person.
Who is building AGI?
OpenAI, Google DeepMind, Anthropic, Meta, xAI and several Chinese labs all say they are working towards AGI or beyond. Safe Superintelligence (SSI) aims directly for safe superintelligence.
Sources and further reading
- Google DeepMind, Levels of AGI: Operationalizing Progress on the Path to AGI (2023)
- OpenAI, OpenAI Charter
- ARC Prize, ARC-AGI benchmark
- Alan Turing, "Computing Machinery and Intelligence", Mind, 1950
Leave a Reply