The three terms side by side
| Narrow AI | AGI | ASI (superintelligence) | |
|---|---|---|---|
| Scope | Specific tasks or domains | Wide range of cognitive tasks, including learning new ones | Virtually all domains of interest |
| Level vs humans | Can be below, equal or far above humans at its task | Roughly human-level on most tasks | Greatly exceeds the best humans |
| Status per sources | Exists | Hypothetical or disputed; no agreed test [2] [1] | Hypothetical [2]; "not yet achieved" in DeepMind's table [1] |
| Example from the sources | AlphaFold, Deep Blue, spell checkers [1] | Definitions differ; see below | A general system that outperforms 100% of humans on a wide range of tasks [1] |
Notice that "narrow" is about breadth, not skill. Google DeepMind's researchers classify AlphaFold as superhuman narrow AI, because it predicts protein structure better than any human but does only that [1]. Bostrom's chess example makes the same point [3].
A more useful picture: levels of performance and generality
Because "AGI" is a fuzzy bucket, a 2023 Google DeepMind paper proposes a grid: five performance levels (Emerging, Competent, Expert, Virtuoso, Superhuman) against two kinds of generality (narrow or general) [1].
| Level | Performance | Narrow example | General |
|---|---|---|---|
| 1 Emerging | Equal to or somewhat better than an unskilled human | Simple rule-based systems | Frontier chat models of 2023 (their rating) |
| 2 Competent | At least 50th percentile of skilled adults | Smart speakers, toxicity detectors | Not achieved at the time of writing |
| 3 Expert | At least 90th percentile of skilled adults | Spelling and grammar checkers, image generators | Not achieved |
| 4 Virtuoso | At least 99th percentile | Deep Blue, AlphaGo | Not achieved |
| 5 Superhuman | Outperforms 100% of humans | AlphaFold, AlphaZero, Stockfish | ASI: not achieved |
The paper's judgements describe systems as of September 2023 and are the authors' estimates; it does not offer a benchmark to settle them [1]. We are not aware of an equivalent official rating for 2026 systems and do not make one up.
Why definitions of AGI keep differing
The same paper reviews nine definitions and finds they disagree about what to measure [1]:
- OpenAI's charter: highly autonomous systems that outperform humans at most economically valuable work.
- Shane Legg and others: a machine able to do the cognitive tasks people can typically do, which raises the questions "which tasks?" and "which people?".
- Mustafa Suleyman's "artificial capable intelligence": complex multi-step tasks in the open world, tested by turning $100,000 into $1,000,000.
- Agüera y Arcas and Norvig: today's frontier language models already are AGI, because they are general; DeepMind's authors reply that generality also needs reliable performance.
When someone says "we have AGI" or "we are years from AGI", ask which definition they use.
From AGI to ASI
IBM calls realizing AGI "a big step" toward developing an ASI and describes ASI as still theoretical [2]. Bostrom argued in 2009 that the step from roughly human-level machine intelligence to superintelligence would likely be much quicker than the step to human level, possibly through self-improvement [4]. That is an argument, not an observation; it has been disputed, and the gap between the two is one of the main uncertainties in the timelines debate covered on What is superintelligence?.
Where does "SI" fit?
In AI writing "SI" often just abbreviates superintelligence, so it sits at the ASI end. In US federal usage since September 2026 it has been defined as the same technologies the law already calls AI, so it sits at the "AI" end. Read SI vs AI before relying on the letters alone [5].