Artificial General Intelligence: Definitions, Evidence Thresholds, and Limits

OpenAI’s charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work.

This is OpenAI’s institutional definition. It is not a consensus standard.

Google DeepMind’s Levels of AGI framework separates breadth or generality from depth or performance and treats autonomy as a deployment characteristic that must be considered alongside capability.

The framework is an ontology for comparing progress, not proof that any current system has reached AGI.

A defensible public claim should specify tasks, breadth, reliability, autonomy, adaptation, resource use, and the human comparison group.

Without those dimensions, the label hides more than it explains.

Breadth should cover multiple cognitive domains and unfamiliar tasks rather than a hand-picked set of benchmark families.

General competence cannot be established by excellence in one domain.

Reliability should be reported at multiple success thresholds, with failures, retries, scaffolding, time, and cost visible.

A system that occasionally succeeds is different from one that can be delegated high-stakes work.

Real-world evidence should include long-horizon tasks with incomplete information, social coordination, changing goals, and consequences that cannot be reduced to an automatic score.

This is stricter than most current public benchmark environments.

METR defines a task-completion time horizon as the human-expert task duration at which an agent is predicted to succeed at a chosen reliability level.

METR says its current suite is primarily software engineering, machine learning, and cybersecurity and explicitly warns that an eight-hour horizon does not mean all jobs can be automated.

ARC-AGI-3 scores completion and action efficiency relative to controlled human baselines across interactive games.

Its Relative Human Action Efficiency is a specific evaluation methodology, not a universal definition of intelligence.

A credible AGI case would require convergent evidence across complementary evaluations rather than one headline score.

Different evaluations expose different slices of performance, reliability, efficiency, and adaptation.

Leading one benchmark does not prove broad, reliable competence across economically valuable work.

The benchmark’s task distribution, contamination controls, scaffolding, and failure modes remain material.

A model’s fluent self-description or claim of consciousness is not an audited capability evaluation.

Language plausibility and demonstrated competence are separate evidence classes.

Large economic impact can arrive before AGI, and an AGI label does not by itself establish broad economic benefit.

Capability, deployment, diffusion, ownership, and distribution must be evaluated separately.

AGI could accelerate research, design, and coordination, but it would remain an input into energy, manufacturing, logistics, law, ownership, and distribution systems.

This distinction is the bridge from capability research to the Universal Basic Abundance framework.