AGI vs ASI: the difference, and the rest of the glossary from narrow AI to frontier models

AGI vs ASI is the question behind most of the arguments about where AI is heading, and the two terms are routinely mixed up. In short: artificial general intelligence, AGI, is a system that can do roughly what a capable person can do across a wide range of thinking tasks. Artificial superintelligence, ASI, is a system that does it better than every person. One is a threshold at the human level, the other is the far side of it. Neither is a settled, measurable thing yet.

The short version

• Narrow AI does a specific task, often better than any human: chess, protein structure prediction, spelling checks.
• AGI is general: it can learn and do many kinds of tasks without being rebuilt for each one.
• ASI is general and better than all humans.

That ladder is how the field usually draws it. The difficulty is that every rung is defined by comparison with people, and people vary. So we need to pin down each term.

AGI

Wikipedia describes AGI as "a hypothetical type of artificial intelligence that matches or surpasses human capabilities across virtually all cognitive tasks." The page contrasts it with narrow AI, which is competent only at specific tasks, and says an AGI can generalize knowledge, transfer skills between domains and solve novel problems without task-specific reprogramming.

Note the word "hypothetical." There is no accepted test, and companies define it differently. That matters when a chief executive says AGI is close, because the claim can be true under one definition and false under another.

ASI

Wikipedia describes artificial superintelligence as "a hypothetical type of AGI that is much more generally intelligent than humans." Nick Bostrom's widely quoted phrasing is "any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest." We unpack that in what is superintelligence.

A scale that puts numbers on it

The most useful attempt to replace vague labels with a scale comes from Google DeepMind researchers, including Meredith Ringel Morris and Shane Legg. Their paper, "Levels of AGI for Operationalizing Progress on the Path to AGI," was first submitted on November 4, 2023, appeared at ICML 2024, and was last revised September 24, 2025. It classifies systems by depth (performance) and breadth (generality). The performance levels, compared with skilled adults:

• Level 1, Emerging: equal to or somewhat better than an unskilled human.
• Level 2, Competent: at least the 50th percentile of skilled adults.
• Level 3, Expert: at least the 90th percentile.
• Level 4, Exceptional: at least the 99th percentile.
• Level 5, Superhuman: outperforms 100% of humans.

For a narrow system, Level 5 is already real: the authors classify AlphaFold as Superhuman Narrow AI. For a general system, Level 5 is ASI. In the paper's table at the time of writing, general systems at Competent, Expert, Exceptional and Superhuman were all "not yet achieved," and the authors said "Competent AGI" best corresponds to many prior conceptions of AGI. The table placed ChatGPT, Bard, Llama 2 and Gemini at the bottom general rung, Emerging AGI. That is a 2023 judgment by a lab that builds these systems, and the paper warns that placements are approximate. It is not a verdict on today's models.

What the scale gets right is that AGI is not one cliff. It is a ladder, and a system can be Expert at some tasks while only Emerging overall.

The rest of the glossary

Frontier models. The most capable general-purpose systems at a given time. The term is used in the White House accord's title ("Frontier Responsibilities") without a legal definition we could find; we cover the pledge in the accord tracker.

Superintelligence. In everyday usage it overlaps with ASI. Companies use it in branding, which is a different thing: Meta's Mark Zuckerberg wrote in July 2025 about "personal superintelligence," and Safe Superintelligence Inc. is a company name.

Powerful AI. Dario Amodei's term, from his January 2026 essay: a model "smarter than a Nobel Prize winner across most relevant fields," able to do tasks that take "hours, days, or weeks" and to run as millions of instances. It resembles AGI more than ASI, although "a country of geniuses in a datacenter" sounds larger.

Super Intelligence, "SI." The US government's term since the September 29, 2026 order, which tells agencies to use it instead of "AI" in non-statutory documents. The legal definition is unchanged: Section 3(a) keeps 15 U.S.C. 9401(3). So "SI" in a federal press release means today's AI. Details are in our explainer on the order.

The singularity. A hypothesized point of runaway change, not a kind of system. It is related to ASI but not the same.

Why the confusion matters

Three habits cause most of the noise. First, people use AGI for whatever they think is next. Second, ASI gets used for any system that beats humans at one thing, which the DeepMind scheme would call narrow. Third, a forecast of "AGI by year X" is quoted without its definition, so the same words carry different dates. For example, the AI Futures Project's "superhuman coder" milestone is explicitly not the full AGI definition used by other forecasters, which is part of why their dates differ; see the AGI timeline tracker.

Our take

Use the DeepMind levels as a checklist: for any claim, ask which level, on which tasks, tested by whom. If the answer is "it feels like AGI," it is a vibe, not a measurement. We would treat ASI as a hypothesis and AGI as a moving target until someone publishes a test that skeptics and boosters both accept.