What Is General Artificial Intelligence? An Evidence Guide to AGI

A terminology note matters here: AGI conventionally expands to artificial general intelligence. “General artificial intelligence” is a common word-order variation and the search phrase used in this page’s original title. Both usually point to the same debate.

Four terms that should not be collapsed into one

The clearest way through the hype is to separate the category of system, the breadth of its uses, the proposed capability threshold and the way it operates.

These definitions answer different questions. The OECD definition helps policy makers decide what counts as an AI system. The 2026 International AI Safety Report discusses the capabilities and risks of today’s most capable general-purpose systems. OpenAI’s Charter states a mission-specific AGI threshold. DeepMind’s framework tries to make progress more measurable by separating breadth from performance. None should be silently substituted for another.

What evidence would make an AGI claim meaningful?

Start with breadth: which materially different tasks and domains are represented? A model completing code, prose and questions through the same text interface may be broadly useful, but the test still needs to show what population of human work those tasks represent.

Then ask about depth: compared with whom, at what performance level and with what reliability? “Human-level” could mean an untrained participant, a typical worker or a domain expert. Average performance can also hide catastrophic failures that matter in a real deployment.

Finally, look for transfer and adaptation. An important test uses unfamiliar, held-out problems and discloses what training, prompting, tools, examples and human intervention were allowed. Otherwise a result may measure preparation for a test rather than the ability to acquire a new skill.

Why one benchmark cannot settle the question

A benchmark is an instrument, not a verdict. It samples particular abilities using a particular interface, dataset, scoring rule and budget. Strong performance can be genuine progress while still leaving other definitions of AGI unresolved.

ARC-AGI is a useful example because its designers deliberately focus on skill acquisition and novel problems rather than accumulated knowledge. The 2026 ARC-AGI-3 technical report introduces interactive environments that require exploration, goal inference and planning. It reports that humans in its study solved all tested environments while evaluated frontier systems scored below one per cent as of March 2026. That is evidence about those systems on that benchmark under its protocol—not a universal measurement of every component of intelligence.

Minimum questions for any benchmark headline

1. Was the task set held out, and how was possible training-data exposure investigated?
2. Which prompts, tools, memory, retries, time and human help were allowed?
3. Is the reported result an average, best run or selected demonstration?
4. What do humans score under comparable time, information and interface conditions?
5. Which failure modes, costs and safety constraints disappear from the headline number?
6. Can independent evaluators reproduce the protocol and challenge the conclusion?

What can be said responsibly in July 2026?

The International AI Safety Report 2026 says general-purpose AI capabilities have continued to improve, including through techniques applied after initial training. It also describes several plausible paths to 2030: progress could slow, continue at recent rates or accelerate. That range is a useful antidote to treating one forecast as settled fact.

This guide therefore does not declare that a particular product is or is not AGI for every possible definition. It makes a narrower claim: no shared label removes the need to define the threshold, publish the test conditions and examine contrary evidence. An organisation may announce that its own threshold has been reached while independent researchers reasonably reject the definition or evidence.

What this means for an Australian builder today

A team does not need an AGI forecast to make a useful AI decision. Start with a real workflow, accountable owner, permitted data, measurable outcome and human intervention point. Test the system that will actually be deployed, including predictable failures, rather than borrowing a frontier-model headline.

The Australian Government National AI Centre’s current foundations guidance recommends accountability, risk management, transparency and human control appropriate to the use. Those controls follow from what the system does and whom it affects—not from whether marketing calls it AI, an agent or AGI.

• Learn the basic distinction first in MLAI’s plain-English guide to AI.
• If the system takes actions, map its boundaries using the guide to AI agents.
• Bring disputed claims and methods to an MLAI event for discussion with the Australian community.
• When you have a defined workflow and success measure, scope the practical system through MLAI Studio.