AGI
AGI, or artificial general intelligence, is a label for a proposed AI system with broad capability across many tasks or domains rather than competence confined to a narrow function. Definitions disagree about the required breadth, performance level, autonomy, learning ability, and economic usefulness. AGI is therefore a research goal and classification problem, not one universally accepted test or a status that follows automatically from success on a particular benchmark.
Origin and context
A 1997 conference paper by Mark Gubrud contains the earliest use reviewed here. Goertzel and Pennachin's 2007 volume then named and organized a research area explicitly focused on engineering general intelligence. Institutional definitions later diverged. OpenAI's 2018 Charter framed AGI around highly autonomous systems outperforming humans at most economically valuable work, whereas a 2023 Google DeepMind preprint separated breadth, performance, and autonomy and proposed levels rather than one binary finish line.
Why it matters
AGI claims influence research priorities, investment, safety programs, governance proposals, and public expectations. Without a stated definition, two organizations can use the same label for materially different capability thresholds, and a prediction about arrival may be impossible to compare with another. A useful assessment names the task distribution, performance reference, reliability, adaptability, autonomy, and deployment conditions. For skills analysis, broad benchmark performance is not the same as dependable execution of real work across contexts, tools, rules, and consequences.
Example
Suppose a model exceeds typical human performance on a broad benchmark suite but cannot reliably learn a new workplace process, operate tools safely, or recognize when to defer. One framework may call it an early or competent level of general AI; another may say it falls short of AGI. The disagreement cannot be resolved by the acronym alone. Reviewers should publish the breadth and depth criteria, compare against an explicit human or system baseline, and report autonomy separately from capability.
How it differs
Superintelligence
Superintelligence describes a hypothetical level far beyond the best human performance across very broad cognitive domains. AGI usually emphasizes generality and some reference level of competence; a system could satisfy a stated AGI definition without being superintelligent.
Jagged Frontier
The jagged frontier describes uneven capability across tasks that may appear similar. It is an empirical warning against inferring generality from selected successes and helps explain why AGI evaluation requires breadth as well as peak performance.
Maturity and evidence
Maturity is rated 4 for the vocabulary and research program, not for the achievement of AGI. The term has documented use across decades and adoption by publishers, research groups, and laboratories. Its referent remains contested: definitions and proposed levels vary, and there is no independent authority that can certify a system against a universally accepted threshold.
Limits and open questions
AGI does not necessarily imply consciousness, personhood, benevolence, embodiment, or unrestricted autonomy. Human-level is also underspecified because humans vary and tasks depend on tools, time, training, and context. Benchmark contamination, selective demonstrations, and rapid model updates can further complicate claims. Any assertion that AGI exists or is near should be read against the speaker's definition, evidence, evaluation access, and incentives, with safety consequences assessed separately from the label.
Related terms
References
- Nanotechnology and International SecurityFifth Foresight Conference on Molecular Nanotechnology / Mark Gubrud · 1997-11 · class A
- Artificial General IntelligenceSpringer / Ben Goertzel and Cassio Pennachin · 2007-01-17 · class A
- Levels of AGI for Operationalizing Progress on the Path to AGIGoogle DeepMind researchers / arXiv · 2023-11-04 · class A
- OpenAI CharterOpenAI · 2018-04-09 · class A
Last updated: 2026-09-07
This term is also covered in the Skills Atlas as model evaluation skill.
This term is also covered in the Skills Atlas as benchmark analysis skill.
This term is also covered in the Skills Atlas as ai risk management skill.