Artificial general intelligence (AGI) is a hypothetical form of artificial intelligence that can understand, learn and perform a broad range of cognitive tasks at or above human level, transferring knowledge between domains instead of excelling only at the tasks it was built or trained for. The older term strong AI is sometimes used as a synonym, although in philosophy it can also refer to the claim that a machine could have a genuine mind.
Among AI fundamentals, AGI is defined mainly by contrast with narrow AI, which covers every deployed system today: chess engines, image classifiers, recommendation models and even large foundation models that show wide but uneven competence. There is no agreed technical definition or test for AGI, and whether or when it will be reached is openly disputed among researchers.
What makes intelligence general
Generality is usually described through a set of capabilities rather than a single benchmark. An AGI system would need broad competence across many task types, the ability to learn new tasks efficiently from limited data or instruction, transfer of knowledge between domains, robust reasoning and planning over long horizons, and the ability to operate in novel situations that were not anticipated in training.
Current systems satisfy some of these properties partly. Large language models perform many tasks from instructions alone, yet they can fail at simple problems outside their training distribution, struggle with reliable long-horizon planning and lack persistent learning after deployment. Experts disagree about whether scaling such models will close these gaps or whether new architectures are required.
The table below summarizes the capabilities most often cited as requirements for general intelligence.
| Capability | Meaning | Status in current systems |
|---|---|---|
| Breadth of tasks | Competence across many unrelated domains | Wide but uneven |
| Sample-efficient learning | Learning new skills from few examples | Partial, mainly through prompting |
| Transfer and abstraction | Applying concepts in new contexts | Inconsistent outside training data |
| Long-horizon planning | Pursuing goals over many steps | Limited reliability |
| Continual learning | Updating knowledge after deployment | Largely absent in deployed models |
| Grounding in the world | Linking symbols to perception and action | Emerging through multimodal and robotic systems |
Competing definitions of AGI
Because the term is used by researchers, companies and policymakers with different goals, its definitions vary substantially. Some focus on matching human performance, some on economic impact and some on the ability to learn rather than on any fixed skill set. Claims that AGI has or has not been reached usually depend on which definition is being applied.
| Definition family | Core criterion | Main criticism |
|---|---|---|
| Human-level performance | Matches typical humans on most cognitive tasks | Unclear which tasks and which humans |
| Economic definition | Can perform most economically valuable work | Depends on markets and deployment, not capability |
| Learning-efficiency definition | Acquires new skills as efficiently as humans | Hard to measure in practice |
| Levels framework | Graded levels of performance and generality | Thresholds remain somewhat arbitrary |
| Behavioral tests | Passes tests such as the Turing test | Tests can be passed by imitation without generality |
The Turing test, proposed in 1950, is the best-known behavioral criterion, but conversational imitation has proven a weak indicator of general capability. Proposed alternatives include broad academic and professional benchmarks, novel abstract-reasoning puzzles designed to resist memorization, and practical tests such as completing open-ended real-world tasks.
AGI compared with narrow AI and related concepts
AGI sits on a spectrum of capability concepts. Distinguishing them prevents a common confusion: broad usefulness is not the same as general intelligence, and general intelligence is not the same as superintelligence.
| Concept | Scope | Status |
|---|---|---|
| Narrow AI | One task or a closely related family of tasks | Widely deployed |
| Foundation model | Broad pretrained model adapted to many tasks | Deployed, competence uneven |
| Artificial general intelligence | Human-level competence across most cognitive tasks | Hypothetical, criteria disputed |
| Artificial superintelligence | Far beyond the best humans in nearly all domains | Speculative |
Machine learning, deep learning and reinforcement learning are methods, not capability levels. Most proposed paths toward AGI combine them, for example large self-supervised models with reinforcement learning for reasoning and tool use, but none of these methods is in itself a definition of general intelligence.
Proposed approaches toward AGI
Research directions differ in what they consider the missing ingredient. The dominant industrial approach scales neural networks, data and compute, on the hypothesis that general capabilities emerge from scale. Other researchers argue for explicit world models, neurosymbolic combinations of learning and logic, embodied learning in physical or simulated environments, or architectures inspired by cognitive science.
- Scaling hypothesis: larger models trained on more data and compute gain broader abilities.
- Agentic systems: models that plan, use tools, keep memory and act over many steps.
- World models: learned internal simulations of how environments respond to actions.
- Neurosymbolic AI: neural perception combined with structured symbolic reasoning.
- Embodied AI: learning through interaction with physical or simulated environments.
Risks, governance and common misconceptions
AGI is a central topic in AI safety and governance, because a system with general capabilities and autonomy could act in ways that are hard to predict or control. The key technical concern is alignment: ensuring that a capable system pursues the goals its developers and society intend, even in situations not covered by training.
| Misconception | Correction |
|---|---|
| Fluent conversation proves general intelligence | Fluency can coexist with brittle reasoning |
| AGI implies consciousness | General capability and subjective experience are separate questions |
| AGI will arrive at a single clear moment | Capabilities more likely grow gradually and unevenly |
| Benchmark records equal general ability | Benchmarks can be saturated without real-world robustness |
In simple terms
Today’s AI tools are specialists. One recognizes faces, another translates text, another writes drafts. Artificial general intelligence would be a generalist: one system that could learn almost any mental task about as well as a capable person.
Think of the difference between a calculator and a new colleague. The calculator is perfect at arithmetic and nothing else. The colleague can learn a new job in a few weeks. No machine has clearly reached that level yet, and experts disagree about when or whether it will.
A practical example
Consider a mid-sized logistics company evaluating AI tools. Its current systems are narrow: one model forecasts parcel volumes, another optimizes delivery routes, and a language model drafts customer replies. Each needed separate data, training and integration, and none can take over the work of another.
A system meeting most AGI definitions would behave differently. Given access to the same tools, it could study the warehouse data, learn the dispatch process and handle an unexpected port closure by replanning routes, contacting customers and adjusting staffing. That difference between assembling many specialists and employing one adaptable generalist is what researchers and planners mean when they discuss AGI.