The three levels of AI are artificial narrow intelligence (ANI), artificial general intelligence (AGI) and artificial superintelligence (ASI). Narrow AI is built for specific tasks, and most AI systems in use today fall into this group. AGI would handle intellectual work about as broadly and flexibly as a person. ASI would outperform the best humans in nearly every cognitive field. AGI has not been shown beyond dispute, although general-purpose models such as ChatGPT make the line harder to draw, and ASI remains hypothetical.
Keep one caveat in mind from the start. This three-way split is a convention, not an official standard, and researchers draw the lines in different places. Google DeepMind, for example, rates AI systems on two separate axes: how broad their skills are and how well they perform compared with people.
In short:
- Narrow AI is everywhere, from spam filters to chess engines, and it can be superhuman at a single task.
- AGI has no universally accepted definition or test, and in September 2026 there is no scientific consensus that any system has reached it.
- ASI would have to beat the best humans broadly, not in one domain. It is hypothetical, and nobody can reliably say when or whether it will arrive.
Narrow AI can reach any level of performance. Only breadth and performance together make AGI or ASI.
ANI vs AGI vs ASI
ANI, AGI and ASI differ in how widely a system can apply its skills and in how it compares with human performance. The table puts them side by side.
| Dimension | ANI | AGI | ASI |
|---|---|---|---|
| Also called | Narrow AI or weak AI | General AI, sometimes strong AI | Superintelligence |
| Task range | One task or a defined set of tasks | A wide range of intellectual tasks | Nearly all cognitive tasks |
| Adaptability | Low outside its training | High, handles new kinds of problems | Beyond human adaptability |
| Knowledge transfer | Limited | Across domains | Across domains, better than humans |
| Performance compared with humans | From below human to superhuman, within its domain | Roughly human level on most cognitive tasks, depending on the definition | Better than the best humans in nearly every field |
| Autonomy | Chosen at deployment | Chosen at deployment, within what capability allows | Chosen at deployment, if it can be controlled |
| Status (September 2026) | Exists and is widely used | Not generally accepted as achieved | Hypothetical |
| Examples | Spam filters, Stockfish, AlphaFold | No undisputed example | None |
| Main limitation | Breaks down outside its domain | No agreed definition or test | Unknown whether it is possible, and hard to control if it is |
Superhuman performance on one task does not make a system ASI. A chess engine like Stockfish can beat every human player alive, yet it cannot summarise a contract or plan a project. It is narrow AI working at the highest level of performance.
What a “level” of AI actually means
A level of AI describes what a system can do, not how it works inside or how much freedom it gets. Four questions usually decide where a system belongs:
- How wide is the range of tasks it can handle?
- Can it transfer knowledge to situations it has never seen?
- Can it adapt and solve problems on its own when conditions change?
- How does it perform compared with skilled humans?
The labels ANI, AGI and ASI bundle all four answers into one word. That is convenient, but it hides the fact that a system can score high on one question and low on another.
Is intelligence the same as consciousness?
Intelligence and consciousness are different things, and having one does not imply the other. Google DeepMind’s paper states that consciousness and sentience are not “a necessary precursor for AGI”.
Level 1: artificial narrow intelligence (ANI)
Artificial narrow intelligence is AI that is designed or trained for one specific task or a limited set of tasks. It is also called narrow AI or weak AI. Most AI in daily use falls into this category.
Characteristics of narrow AI
Narrow AI has four typical characteristics:
- It works within the limits of its training or design.
- It can outperform humans in its own domain.
- It has no general, human-like range of skills.
- It can fail badly on inputs unlike its training data, and it may not signal that it is failing.
Are ChatGPT and other language models narrow AI?
Current language models sit in a grey zone between narrow and general AI. They are far broader than classic single-task AI, because the same model writes, translates, codes, reads images and uses tools. DeepMind’s framework, first written in 2023, puts ChatGPT, Bard, Llama 2 and Gemini at its lowest level of general AI, which it calls Emerging AGI.
Breadth alone does not make a system AGI, though. There is no accepted test or threshold for AGI, and current models are not widely regarded as proven AGI. Calling ChatGPT “just a chatbot” undersells what it does, and calling it AGI claims more than the evidence shows. Our overview of the top 10 AI chatbots compares what today’s assistants can do.
Level 2: artificial general intelligence (AGI)
Artificial general intelligence is a hypothetical AI system with roughly human-level ability across a wide range of intellectual tasks. It would apply its knowledge and skills broadly and flexibly. The key word is general. The same system should be able to take on a new kind of problem without being rebuilt for it.
Organisations disagree on where exactly the line for AGI sits. The OpenAI Charter describes AGI as “highly autonomous systems that outperform humans at most economically valuable work”. The ARC Prize Foundation defines it as a system’s ability to acquire any skill a human can, as efficiently as a human can. Anthropic CEO Dario Amodei prefers the term “powerful AI” and writes that he dislikes “AGI”. DeepMind calls its own Competent AGI level “probably the best catch-all” for many existing definitions.
Has AGI been reached yet?
There is no broad scientific consensus that AGI has been reached. In 2023, Microsoft researchers argued that GPT-4 “could reasonably be viewed as an early (yet still incomplete) version” of AGI. Others point to gaps that benchmark scores miss. The International AI Safety Report 2026, published in February, describes capabilities as “jagged”. Leading systems may excel at some difficult tasks while failing at simpler ones. Examples include counting objects in an image and recovering from basic errors in longer workflows.
How AGI could be measured
You have to measure AGI along several dimensions at once, not with a single score. The main candidates are the breadth of tasks, the level compared with humans, performance in new situations, robustness and the ability to learn new skills unaided.
Google DeepMind turned the first two into a grid. Its Levels of AGI framework rates a system on generality (narrow or general) and on performance, which it splits into five levels above “no AI”. The table reproduces the paper’s own examples and assessments. They date from the paper’s early versions and are unchanged in the September 2025 revision. So the “not yet achieved” cells are the paper’s view, not a verdict for September 2026.
| Level | Performance compared with humans | Narrow AI example | General AI example |
|---|---|---|---|
| 1. Emerging | Equal to or somewhat better than an unskilled human | Rule-based systems such as SHRDLU | ChatGPT, Bard, Llama 2, Gemini |
| 2. Competent | At least the 50th percentile of skilled adults | Smart speakers, IBM Watson | Not yet achieved |
| 3. Expert | At least the 90th percentile of skilled adults | Grammarly, DALL-E 2 | Not yet achieved |
| 4. Exceptional | At least the 99th percentile of skilled adults | Deep Blue, AlphaGo | Not yet achieved |
| 5. Superhuman | Outperforms 100 percent of humans | AlphaFold, AlphaZero, Stockfish | ASI, not yet achieved |
Level 3: artificial superintelligence (ASI)
Artificial superintelligence describes a hypothetical AI that would consistently beat the best humans at nearly every cognitive task. Philosopher Nick Bostrom gave the classic definition in 1997. He described an intellect “much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills”.
How does ASI differ from AGI?
AGI would roughly match broad human intelligence, while ASI would far exceed it. A hypothetical ASI could outperform the best human experts in fields such as:
- scientific research;
- strategy and planning;
- software development;
- creative problem solving;
- medical analysis;
- technological innovation.
What is superintelligence?
Superintelligence now has two meanings, and they are far apart. In research, it is an intellect far smarter than the best human minds in practically every field. In US government language, it has been the official name for all AI since 29 September 2026.
The research meaning is described an intellect much smarter than the best human brains in practically every field. No system meets that bar today. ASI, or artificial superintelligence, is that idea applied to AI, and it is what this article means by level 3.
The government meaning comes from President Trump. At the UN General Assembly on 22 September 2026, he said that “the use of the word artificial makes intelligence sound fake”. He announced that US documents would say super intelligence instead. A week later, he signed an executive order that tells federal agencies to use “Super Intelligence” and “SI” in place of “Artificial Intelligence” and “AI”.
Which companies say they are building ASI?
Several AI companies now name superintelligence as their goal, although ASI itself remains hypothetical. Safe Superintelligence Inc., led by Ilya Sutskever, says it has one goal and one product: a safe superintelligence. OpenAI, Meta, with its Meta Superintelligence Labs, and Microsoft, with its MAI Superintelligence Team, have announced similar aims. These are statements of intent, not evidence of capability.
Why the boundaries between AI levels are so hard to draw
Intelligence itself has no agreed definition, so every line between the levels is open to debate. Shane Legg, who later co-founded DeepMind, and Marcus Hutter collected 70-odd definitions of intelligence back in 2007. Several other problems come on top of that:
- Performance varies sharply by task. A model can be expert at writing code and poor at counting the objects in a picture.
- Fluent answers can make a system seem to understand more than it does. That is also why AI hallucinations are so convincing.
- A system can handle many tasks and still fail unpredictably on some of them, so breadth says little about reliability.
- DeepMind rates what a system could do, not how widely it is used, and a result in the lab is not yet a working product.
Where AI stands in September 2026
In September 2026, general-purpose models are broad but uneven, and no system is generally accepted as AGI. This section collects the figures that change fastest, so check the date before you quote them.
What the latest benchmark results show
Benchmark scores jumped in 2026, but the organisation behind the best-known test still does not call its top model AGI. That test is ARC-AGI-3, a set of puzzle environments that humans can solve in full. In May 2026, GPT-5.5 scored 0.43 percent on the test’s semi-private set with the ARC Prize Foundation’s Standard harness. Four months later, OpenAI’s GPT-6 Astra scored 62.7 percent on the same set and harness, at its maximum reasoning setting. With the Provider Adapter harness, which lets the model use OpenAI’s own context-management features, Astra scored 99.9 percent at its high setting.
These are scores on one benchmark, not a percentage of general intelligence. The ARC Prize Foundation wrote of Astra: “we are not claiming that it is AGI”. OpenAI’s launch post does not describe Astra as AGI either.
What do forecasts say about when AGI will arrive?
Forecasts vary widely, and none of them is a scheduled date. A survey of 2,778 AI researchers appeared as a preprint in 2024 and in a journal in 2025. Its aggregate forecast put a 50 percent chance on unaided machines outperforming humans at every possible task by 2047. That forecast assumes science continues undisrupted, and it uses the survey’s own definition rather than an agreed definition of AGI.
In October 2024, Dario Amodei wrote that what he calls powerful AI could come as early as 2026, “though there are also ways it could take much longer”. His bar is higher than most of the AGI definitions above: a model “smarter than a Nobel Prize winner across most relevant fields”.
How you can judge which level an AI tool has reached
You judge an AI tool by what it does on your own tasks, not by the label it is sold under. Five questions cover most of what matters:
- How many different kinds of task does the system handle without being rebuilt?
- How does it score against skilled humans on each of those tasks, not only on its best ones?
- Does it hold up on problems that cannot have been in its training data?
- Can it learn a new skill on its own, and how much practice does it need?
- Does it stay reliable over long, messy workflows in the real world?
A system that does well on the first two questions but poorly on the last three is better described as a broad tool than as AGI.
For a business, the practical version comes down to task range, reliability and the oversight a tool needs. Three hypothetical examples, not client cases, show how that works:
| Tool (hypothetical example) | Task range | Reliability to test | Oversight needed |
|---|---|---|---|
| An agent that processes supplier invoices | Narrow: one workflow with known document types | Accuracy on a labelled sample of your own invoices, including unusual layouts | Human approval above a set amount, and a log of every action |
| A general-purpose assistant such as ChatGPT or Copilot for staff | Broad: writing, analysis, research and code | Uneven by task, so test the tasks your staff actually use it for, including checks for invented facts | Staff review every output before it leaves the organisation, backed by a clear usage policy |
| A customer service chatbot on top of your knowledge base | Bounded by the documents it can search | Whether answers match the source documents, and what it does when the answer is missing | Handover to a person for complex questions, and regular review of conversations |
None of these tools needs to be AGI to do its job. What matters is whether the task range, reliability and oversight fit the job.
Frequently asked questions
Is generative AI a separate level?
Generative AI describes what a system produces, such as text, images, audio or code, and says nothing about its level. An image generator is narrow AI that happens to create things, not a fourth level between ANI and AGI.
Where do machine learning and deep learning fit?
Machine learning and deep learning are methods for building AI systems. They are not levels above or below ANI, AGI or ASI. A narrow spam filter and a large language model can both be built with deep learning.
Does an AI agent count as AGI?
Being an agent does not make a system AGI. The word agent means that a system can pursue a goal and take actions, such as calling tools or clicking through a website. An agent built for one workflow, such as processing invoices, is a narrow application even when a general-purpose model runs inside it.
What about reactive machines, limited memory and theory of mind?
Reactive machines, limited memory and theory of mind come from a different classification. In a 2016 article, Michigan State University researcher Arend Hintze sorted AI into four types by function, with self-awareness as the fourth. That scheme is about memory and assumed mental abilities, not about breadth and performance. So do not map it one-to-one onto ANI, AGI and ASI.
Is general-purpose AI in the EU AI Act the same as AGI?
The EU AI Act’s term general-purpose AI model covers today’s large models, which can perform a wide range of distinct tasks. That is breadth, not human-level intelligence. Providers have had obligations since 2 August 2025, and models already on the market before then have until 2 August 2027. Our EU AI Act checklist explains what applies to your organisation.
If you plan to use AI in your organisation, an AI assessment shows which of your processes today’s narrow and general-purpose AI can realistically take on. A workshop helps your team separate what these systems do today from what the headlines promise.
