What is AGI?
AGI (Artificial General Intelligence) is a contested concept for an AI system with broad, transferable capability across many cognitive tasks and environments; no single definition, benchmark, or recognized certification currently establishes that a system is AGI.
Quick Facts
| Full Name | Artificial General Intelligence |
|---|---|
| Specification | Official Specification |
How It Works
Definitions and Claim Boundaries
AGI has no universally accepted specification. A useful claim names the reference population, task distribution, performance threshold, allowed tools and scaffolding, learning budget, and deployment conditions. Phrases such as human-level, generally smarter, or able to do most work remain ambiguous without that contract.
Narrow and general intelligence are not separated by whether a system uses an LLM or agent loop. A system can exceed experts on one benchmark and still fail to transfer to unfamiliar tasks. Conversely, broad product coverage can come from routing many specialized components rather than one generally capable model.
Performance, Generality, and Autonomy
The Levels of AGI framework separates depth of performance from breadth of generality and treats autonomy as a deployment dimension. This avoids declaring AGI from a single average score or assuming that a capable model is automatically an autonomous system.
A capability profile should report performance by domain, reliability, novelty, learning efficiency, and the human or system baseline. An autonomy profile should separately state who chooses goals, supplies context, approves actions, monitors progress, and can stop or recover the system.
Measurement Challenges
Benchmark results can reflect training-data exposure, prompt and tool scaffolding, repeated attempts, evaluator leakage, or task-specific optimization. Static suites also underrepresent adaptation to new rules, long-horizon work, social context, physical interaction, and recovery after failure. Skill on known tasks is therefore not the same as efficient acquisition of new skills.
Credible evaluation uses private or fresh tasks, explicit priors and experience budgets, contamination checks, human baselines, confidence intervals, and cross-domain failure analysis. No current benchmark covers every relevant form of generality.
Risk and Governance Without a Binary Label
Operational risk depends on the combination of capability, autonomy, access, scale, and recoverability. A moderately capable system with broad credentials and unattended execution can be more dangerous in one deployment than a stronger model confined to a read-only tool.
Organizations do not need to settle the AGI debate before applying controls. They can measure concrete capabilities, restrict authority, stage evaluations, monitor consequential outcomes, require human approval, and update safeguards as evidence changes. Marketing claims or one benchmark score should not determine deployment permissions.
Key Characteristics
- Definition-dependent: every AGI claim needs an explicit capability and evaluation contract
- Broad rather than merely strong: performance must cover a meaningful distribution of cognitive tasks
- Transfer-sensitive: unfamiliar tasks and efficient skill acquisition matter beyond memorized benchmark skill
- Profile-based: domain results and failure slices should remain visible instead of collapsing into one average
- Autonomy-separable: capability does not by itself determine goals, permissions, or deployment control
- Uncertified: no current standard body or single benchmark grants a generally accepted AGI status
Common Use Cases
- Comparing research definitions of general intelligence without treating them as equivalent
- Designing a multi-domain evaluation that reports breadth, depth, novelty, and learning budget
- Separating model capability claims from the autonomy and permissions of a deployed agent
- Reviewing whether a benchmark result depends on data exposure, tools, retries, or custom scaffolding
- Planning staged governance controls without waiting for a binary declaration that AGI has arrived
Example
Loading code...Frequently Asked Questions
Is there an agreed definition or test for AGI?
No. Research definitions emphasize different combinations of breadth, performance, transfer, learning efficiency, autonomy, or economic work. Frameworks such as Levels of AGI make those dimensions more explicit, but they are proposals rather than a universal certification standard. Any claim should state its chosen definition and evidence.
Does passing a difficult benchmark prove that a system is AGI?
No. A benchmark measures performance under a particular task set, prompt, tool, sampling budget, and scoring rule. High performance can coexist with failures on unfamiliar tasks or different domains. An AGI claim would need broad, contamination-aware, reproducible evidence and clear treatment of adaptation and reliability.
Is autonomy required for AGI?
Definitions differ, but capability and autonomy should be measured separately. A broadly capable model may be deployed as a tightly supervised tool, while a narrower model can be placed in a highly autonomous runtime. Autonomy changes exposure and risk; it does not automatically prove broader intelligence.
Have current AI systems achieved AGI?
There is no generally accepted finding that settles this question because the definition and evaluation standard are disputed. Current systems show strong but uneven capabilities, and public results depend on models, tools, scaffolding, and test conditions. The defensible approach is to report a capability profile rather than a binary label.
How is AGI different from artificial superintelligence?
AGI usually refers to broad capability around a stated human reference level. Artificial superintelligence refers to capability far beyond that reference across a broad range of domains. Both terms remain definition-dependent, and neither label alone specifies autonomy, access, safety, or the controls required for deployment.