AGI vs ASI vs ANI is the classic three rung ladder for AI capability. Artificial narrow intelligence (ANI) is AI built for one task or a tight family of tasks. Artificial general intelligence (AGI) is the research goal of matching people across a wide range of cognitive work. Artificial superintelligence (ASI) would go far past the best humans across nearly every domain that matters. As of November 7, 2026, ANI is everywhere, AGI is disputed with no formal lab declaration, and ASI remains hypothetical. For the pairwise AGI and ASI debate in more depth, see superintelligence vs AGI.
This guide defines each level with lab and textbook sources, shows how Google DeepMind's Levels of AGI matrix maps onto the same idea, and places today's assistants on the ladder without treating marketing claims as settled science.
AGI vs ASI vs ANI at a glance
| Item | ANI (narrow AI) | AGI | ASI (superintelligence) |
|---|---|---|---|
| Core idea | Strong at one task or a limited set | Broadly human level across many cognitive tasks | Far above humans across nearly all domains |
| Also called | Narrow AI, specialised AI, weak AI | General AI (sometimes "strong AI") | Artificial superintelligence |
| Exists today? | Yes, widely deployed | Disputed; no major lab has formally declared it | No, by research definitions |
| Performance vs humans | Can be below, at, or above human level inside its niche | Roughly human level on most cognitive work (definition varies) | Beats essentially all humans on a wide range of tasks |
| Classic examples | Spam filters, chess engines, AlphaFold, speech recognition | None agreed; frontier chatbots are the main debate | None |
| Classic sources | Ada Computer Science; DeepMind "Narrow" column | OpenAI Charter; DeepMind Competent AGI cell | Bostrom Superintelligence; DeepMind top right cell (ASI) |
The short version: "narrow" is about breadth, not about being weak. A system can be superhuman at protein folding and still be ANI. AGI is about breadth at roughly human depth. ASI is breadth and depth far past people.
What ANI (artificial narrow intelligence) means
Ada Computer Science, part of the University of Cambridge outreach programme for computing education, defines artificial narrow intelligence as systems "developed to perform a particular task or a limited range of related tasks." Their ability in one area does not imply equal ability elsewhere. Ada is careful on a point that confuses many readers: "Narrow" describes scope, not skill. A specialised system may outperform people at its particular task.
Everyday ANI includes spam filters, facial recognition, grammar checkers, recommendation rankings and game engines. Deep Blue played chess. AlphaFold predicts protein structures at a level researchers treat as above human experts for that scientific task. Neither system can freely switch to writing a legal brief, tutoring a child in history and debugging an unfamiliar codebase the way a person can.
Modern assistants blur the line because one product can chat, code, summarise and generate images. Ada notes that performing many tasks does not by itself prove AGI; that depends on the definition and the evidence. IBM's overview of AI types still groups today's deployed systems, including text chat products, under narrow or "weak" AI when the framing is "only this category exists in production." Lab papers often use a finer grid (next section). Both views can be true: the product is a bundle of specialised skills, while the capability debate asks whether the bundle has crossed into general human level competence.
What AGI means
There is no single legal or scientific definition. Three framings dominate:
OpenAI's economic definition. The OpenAI Charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work." In Planning for AGI and beyond, OpenAI also describes AGI as "AI systems that are generally smarter than humans," and treats AGI as a point on a continuum that continues toward systems beyond it.
Google DeepMind's ladder. In the ICML 2024 position paper Levels of AGI for Operationalizing Progress on the Path to AGI (Morris et al.; also on the PMLR proceedings page), researchers at Google DeepMind rate systems on performance (Emerging, Competent, Expert, Exceptional, Superhuman) crossed with Narrow versus General. "Competent AGI" (at least median skilled adult performance on a wide range of non physical tasks, including learning new skills) is the cell that matches many everyday "human level AGI" intuitions. The paper said Competent AGI had not been achieved when it was written; frontier chat models were classed as Emerging AGI.
Everyday usage. In ordinary speech, AGI means AI that can do almost any thinking task a person can. That looser meaning is why headlines about whether AGI has been achieved stay contested even when benchmarks jump.
AGI is a capability goal, not a product architecture. An AGI vs LLM comparison makes that split clear: an LLM is a kind of model; AGI is a threshold people argue about.
What ASI means
Artificial superintelligence means an intellect that greatly exceeds the best human minds across virtually all domains of interest. That research sense is the one Nick Bostrom popularised and that safety organisations such as the Center for AI Safety discuss when they talk about extinction level risk from future systems.
DeepMind's Levels of AGI matrix places ASI in the top right cell: Superhuman performance and General breadth. AlphaFold sits in Superhuman Narrow AI, not ASI, because it dominates one scientific task rather than a wide cognitive range. Winning a video game or drafting a novel would not, on Ada's account, prove ASI either. The system would need to beat highly capable humans across many different areas.
OpenAI's AGI planning essay draws the same sequence: carefully build and deploy AGI, then manage the jump to systems beyond it. Whether that jump is possible at all is a separate debate; see is superintelligence possible?.
How DeepMind's matrix maps onto ANI, AGI and ASI
The textbook three letter labels collapse a two axis grid into one ladder. Mapping them reduces confusion:
| Textbook label | Closest DeepMind cells | Why |
|---|---|---|
| ANI | Any Narrow column cell (Emerging through Superhuman) | Breadth is limited even if skill is world class |
| AGI (loose / "emerging") | Emerging AGI | ChatGPT era models: general products, uneven skill |
| AGI (classic "human level") | Competent AGI (and higher General cells short of ASI) | Median skilled adult performance across most cognitive tasks |
| ASI | Artificial Superintelligence (Superhuman × General) | Outperforms 100% of humans across a wide range |
So "AGI vs ASI vs ANI" and "Levels of AGI" are compatible. The three letter version is a coarse public shorthand. The matrix is what researchers use when they need to say that a system can be Superhuman Narrow without being ASI.
AGI vs ASI timeline and the singularity
People searching "agi vs asi timeline" usually want ordering, not a calendar date nobody can prove:
- ANI: specialised systems (today's default).
- AGI: broad human level (or better on OpenAI's economic wording). Disputed for current flagships.
- ASI: far above humans across nearly all domains. Not claimed by any major lab as of November 2026.
- Singularity talk: speculative claims that ASI could trigger very fast further gains, for example via AI systems that improve AI. That is a hypothesis about dynamics, not a fourth capability label. Our AI safety explainer covers why labs and governments care about that path.
Independent evaluators such as METR focus on measured dangerous capabilities and reliability rather than hanging a single AGI or ASI badge on a model. Our guide on how autonomous frontier AI models are summarises what recent system cards say about self improvement thresholds.
No honest source can give a reliable year for AGI or ASI. Lab forecasts move, definitions move, and earlier public timelines have slipped. Treat any specific year as a bet, not a fact.
Where today's systems sit (November 2026)
| System class | Best fit on the three level ladder | Notes |
|---|---|---|
| Spam filters, recommenders, most vision classifiers | ANI | Clear single task scope |
| AlphaFold, top chess or Go engines | ANI (often superhuman in niche) | DeepMind's Superhuman Narrow examples |
| ChatGPT, Claude, Gemini and similar assistants | Debated: ANI product stack vs Emerging AGI | DeepMind paper: Emerging AGI; some vendor marketing and press argue further |
| Formally declared AGI or ASI | None | No major lab declaration as of this writing |
Practical takeaway for users: today's flagship assistants are general products with uneven skills. Excellent at some writing and coding tasks, unreliable at others, and still supervised. That pattern fits DeepMind's Emerging AGI idea more than Bostrom style ASI, and it still lives mostly inside specialised toolchains when you look under the hood.
Pros and cons of the three level framing
Useful when
- You need a shared vocabulary for news, policy briefs or classroom teaching.
- You want to stop people calling every chatbot "superintelligence."
- You need to separate "superhuman at one task" from "superintelligent in general."
Misleading when
- Speakers treat the three labels as official ISO style standards (they are not).
- A single benchmark spike is sold as proof that AGI arrived.
- Federal "SI" (the US rename of ordinary AI policy language) is confused with research ASI. See what SI means instead of AI.
Compared with DeepMind's matrix
- Pros of ANI/AGI/ASI: short, memorable, matches what people type into search.
- Cons: hides the performance axis, so AlphaFold and a weak spam filter both land in "ANI" despite huge skill gaps.
How to keep the terms straight
- Ask breadth: one task, many tasks, or virtually all domains?
- Ask degree: below human, median human, expert human, or beyond every human?
- Ask whose definition: OpenAI economic, DeepMind level, Ada textbook, or federal SI.
- Separate marketing claims from formal lab declarations and independent evaluations.
- For pairwise AGI versus ASI detail, use superintelligence vs AGI; for whether ASI can be built at all, use is superintelligence possible?.
FAQ
What is the difference between AGI, ASI and ANI?
ANI is specialised AI for one task or a limited set. AGI is the goal of roughly human level ability across a wide range of cognitive tasks. ASI (artificial superintelligence) would substantially exceed humans across nearly all domains. ANI exists today; AGI is disputed; ASI is hypothetical.
Is AGI the same as ASI?
No. ASI means artificial superintelligence. On the usual ladder, AGI stops at roughly human level generality; ASI sits above it. DeepMind's matrix puts ASI only in the Superhuman × General cell.
Is ChatGPT ANI or AGI?
It depends who you ask. Textbook and vendor "types of AI" pages often still call today's chat products narrow or weak AI because they are limited compared with a full human. DeepMind's Levels of AGI paper classed ChatGPT era models as Emerging AGI: general products at an early performance level, not Competent AGI. No major lab has issued a formal AGI declaration for any current system as of November 7, 2026. See has AGI been achieved?.
What does AGI vs ASI vs singularity mean?
ANI, AGI and ASI are capability labels. "Singularity" usually refers to a speculative story about very fast capability growth after powerful AI arrives, not a separate rung on the same ladder. You can accept the three levels and still reject singularity timelines.
Has ASI been achieved?
Not by research definitions used by DeepMind, Ada's ASI description or major lab public materials. Narrow systems can already be superhuman at single tasks without counting as ASI.
Is ANI the same as weak AI?
Often yes in older writing. "Weak AI" was John Searle's contrast with "strong AI" (machines that literally understand). In modern product talk, "weak AI" and "narrow AI" are usually synonyms for specialised systems. Prefer ANI or "narrow AI" when you mean limited task scope, because "weak" sounds like low performance and that is not what the term means.



