Artificial general intelligence (AGI) is a hypothetical AI system that can match or exceed human ability across a wide range of cognitive tasks, including ones it was not built or trained for. There is no agreed definition or test for it: researchers and companies define the threshold differently, and they disagree on whether any current system is close.
The contrast is with narrow AI: a system that does one kind of task well, such as playing chess, ranking search results, or resolving a support request. Every AI system deployed in business today, including the most capable large language models, is evaluated against specific tasks.
This page reports how the main definitions differ, attributes each to its source, and makes no prediction about when or whether AGI will arrive.
Whose definition is whose
The definitions fall into a few families: some describe abilities, some describe economic impact, and some describe how efficiently a system learns. They imply very different thresholds.
| Source | Definition, in short | What it measures |
|---|---|---|
| Mark Gubrud, 1997 (first known use of the term) | AI that rivals or surpasses the human brain in complexity and speed and is usable in almost any role where human intelligence would be needed2 | Breadth of use, compared with the human brain |
| Shane Legg and Ben Goertzel (popularized the term in 2001) | A machine able to do the cognitive tasks that people can typically do2 | Human-level performance on non-physical tasks |
| Legg and Hutter, 2007 | General intelligence as an agent's ability to achieve goals in a wide range of environments3 | Performance across environments |
| OpenAI Charter | Highly autonomous systems that outperform humans at most economically valuable work1 | Economic output relative to people |
| François Chollet, 2019 | Intelligence as skill-acquisition efficiency: how quickly a system learns new tasks from limited experience4 | Learning efficiency, not stored skill |
| Morris et al., Google DeepMind, 2023 | A matrix of performance levels and generality, from "Emerging" to "Superhuman"2 | Depth and breadth of capability |
Summaries, not full quotations. Each source sets its own scope and caveats.
AGI as levels, not a single line
Researchers at Google DeepMind proposed rating systems on two axes: how well they perform, compared with skilled adults, and how general they are. The levels below are their performance scale for general systems.2
Emerging
Equal to or somewhat better than an unskilled humanThe authors placed frontier chat models of 2023, such as ChatGPT, Gemini, and Llama 2, here for general tasks.
Competent
At least 50th percentile of skilled adultsListed in the paper as not yet achieved for general systems. The authors call this the best catch-all for many existing AGI definitions.
Expert
At least 90th percentile of skilled adultsNot yet achieved for general systems, per the paper.
Exceptional
At least 99th percentile of skilled adultsThe paper maps OpenAI's labor-replacement definition to roughly this level.
Superhuman
Outperforms all humansCalled artificial superintelligence when general. Some narrow systems, such as AlphaFold, already sit here for one task.
AGI and the systems companies deploy today
Much of the confusion around AGI comes from treating capable products as general ones. The table separates the terms.
| Term | What it names | Exists today | In a support operation |
|---|---|---|---|
| Narrow AI | A system built and tested for a defined set of tasks | Yes | An intent classifier, a fraud model |
| Large language model | A general-purpose text model, broad but uneven | Yes | The model behind a support agent |
| AI agent | A system that pursues a goal with tools and permissions | Yes | An agent that processes a return end to end |
| AGI | Human-level or better ability across most cognitive tasks | Disputed; no agreed test | Not a product category |
| Superintelligence | General ability that outperforms all humans | No | Not applicable |
What narrow agents do instead
A company does not need general intelligence to automate most customer contacts. It needs a system that is reliable on a known set of tasks, with defined data, tools, and limits. These are narrow by design, and that is what makes them testable.
One domain, done well
A retailer's support agent handles order status, returns, and delivery problems, and hands everything else to a person.
known intents → tools → resolve or hand overAdvice from one catalog
A shopping agent recommends only from the company's products and stock. Its knowledge stops at the catalog, on purpose.
needs → catalog → recommendPolicy-bound actions
An installment provider's agent can reschedule a payment within set limits and must escalate anything outside them.
request → check limits → act or escalateA defined process
A procurement agent matches invoices to orders. It is measured on match accuracy, not on general reasoning.
invoice → match → post or flagIllustrative examples. Narrow scope is a design choice that makes an agent measurable and safe to deploy.
How the AGI debate affects buying decisions
The debate matters to a business mainly through expectations. These are the common traps.
Waiting for a general system
ControlJudge AI on the tasks you have now. A narrow agent that resolves a defined share of contacts can be measured in a proof of concept; a future general system cannot.
Treating breadth as reliability
ControlA model that can discuss anything still fails on specific policies and data. Test on your own conversations, with your own definitions of a correct answer.
Vendor claims
ControlAsk which definition of AGI a vendor uses and what evidence supports it. No benchmark is accepted as an AGI test.
Over-scoping permissions
ControlCapable models tempt teams to grant broad access. Give each agent the narrowest permissions its tasks need, with guardrails and human approval for risky actions.
Common questions
Does AGI exist yet?
There is no consensus. Some researchers argue that current models show early signs of general intelligence; others argue they do not meet any rigorous definition. No agreed test exists to settle it.
Is ChatGPT AGI?
By most published definitions, no. A 2023 framework from Google DeepMind placed ChatGPT at "Emerging AGI," its lowest level, while OpenAI's own charter defines AGI as systems that outperform humans at most economically valuable work.
What is the difference between AGI and AI agents?
An AI agent is a system that pursues a specific goal with tools and permissions, such as resolving a return. AGI refers to general, human-level ability across most tasks. Agents in production today are narrow.
What is the difference between AGI and superintelligence?
AGI usually means human-level ability across most tasks. Superintelligence means general ability that outperforms all humans. In the DeepMind levels, general superintelligence sits above all AGI levels.
References
- OpenAI. OpenAI Charter. 2018.
- Morris, M. R. et al. Levels of AGI for Operationalizing Progress on the Path to AGI. Proceedings of ICML 2024 (position paper).
- Legg, S. and Hutter, M. Universal Intelligence: A Definition of Machine Intelligence. Minds and Machines 17(4), 2007: 391-444.
- Chollet, F. On the Measure of Intelligence. arXiv, 2019.
- Bubeck, S. et al. Sparks of Artificial General Intelligence: Early experiments with GPT-4. Microsoft Research, arXiv, 2023.