AI vs AGI vs SI vs ASI: Differences, Definitions and Examples

The four terms answer three different questions, and most confusion comes from mixing them up. Artificial intelligence (AI) is the broad field: any machine doing something that normally takes intelligence. Artificial general intelligence (AGI) is a question of breadth: can one system learn and handle many kinds of problems? Superintelligence (SI) is a question of level: does the intellect far exceed the best humans? Artificial superintelligence (ASI) is a question of substrate: it is superintelligence built from machines.

Our take: treat ASI as a subset of SI, not a fourth step on a ladder. The usual "AI, then AGI, then SI, then ASI" staircase makes SI and ASI sound like different levels. They are not.

What do AI, AGI, SI and ASI each mean?#

The difference between AI and AGI#

AI. The field of building machines that do tasks which normally need human intelligence. Almost everything deployed today is AI, from spam filters to chatbots.

AGI. A system with general, human-level (or beyond) ability to learn, reason and apply knowledge across many tasks, in Stanford HAI's wording. The definition is contested; see the next sections.

What is artificial superintelligence? SI and ASI#

SI. Nick Bostrom, who gave the term its modern meaning, describes superintelligence as an intellect much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills (How Long Before Superintelligence?, 1998). The definition does not say what the intellect is made of.

ASI. Superintelligence implemented in machines. Google DeepMind's 2026 paper defines its version of ASI more narrowly: a system exceeding the cognitive capacity of large human organisations.

Is AGI the same as superintelligence?#

No, and the cleanest way to see it is to separate breadth from level.

AlphaFold is narrow but superhuman at one job. AlphaFold predicts protein structures from amino-acid sequences, work recognised by the 2024 Nobel Prize in Chemistry (awarded to Demis Hassabis and John Jumper of Google DeepMind, with David Baker, per DeepMind's announcement of 2024-10-09). It cannot write an email. It is narrow AI, not AGI, whatever its level on proteins.

An AGI could be the reverse: broad, but only about as good as a skilled person. Google DeepMind's "Levels of AGI" framework (Morris et al., first submitted 2023-11-04) makes this explicit. It scores performance (from Emerging, roughly an unskilled human, up to Superhuman, beating 100% of humans), generality (narrow versus a wide range of non-physical tasks, including learning new skills) and autonomy as separate dimensions. A system can sit high on one and low on another.

Why is there no agreed definition of AGI?#

Stanford HAI says people mean different things by "human-level intelligence" and that there is no universally accepted test, so claims are hard to verify. In practice, three approaches compete:

1. Capability breadth. Can the system do a sufficiently wide range of cognitive tasks?
2. Human equivalence. Does it match human performance across the domains humans handle?
3. General learning. Can it pick up unfamiliar skills efficiently, with little training?

Many popular definitions add requirements, such as lifelong learning without retraining. These are proposed criteria, not agreed prerequisites. Our take: when a company or headline says "AGI achieved", ask which of the three it means and who ran the test.

What makes an intelligence general, and what makes it super?#

Our own working checklist, not a standard. A system is general to the extent that it shows:

• reasoning and inference
• learning new skills
• transferring knowledge between domains
• planning over long horizons
• memory and accumulated experience
• adapting to unfamiliar situations
• checking and correcting its own work

A system is super to the extent that it exceeds the best humans on dimensions such as problem-solving quality, speed, scientific and technological creativity, strategic reasoning, and the ability to run many copies in parallel. Speed and parallelism do not by themselves make an intellect better, only faster or more numerous. That is why we keep level and breadth apart.

Intelligence, autonomy and consciousness are three different properties. A system can act autonomously without being very capable, and can be very capable with no subjective experience. Nothing in the AGI or ASI definitions requires consciousness.

Does AGI exist today?#

Not in a way anyone can verify. With no accepted test, the answer depends on which definition you pick.

What is documented: the 2023 Levels of AGI paper placed ChatGPT, Bard and Llama 2 at "Emerging AGI", a Level 1 general system, while noting competent performance on some tasks. Frontier models have moved on since, and we did not find an independent, widely accepted evaluation that places any of them at human-level generality. We could not confirm any claim that a current system is AGI.

Do not file today's general-purpose models under "narrow AI" either. A model that writes code, reads images and drafts contracts is broad in a way AlphaFold is not. It is also unreliable in ways a human generalist is not. Our label: broad, uneven, and not yet measured against a settled bar.

Current systems also do not simply freeze after training. Deployed models usually have fixed weights, but they can adapt through context, external memory and tools, and developers retrain them in later versions. These are different mechanisms from a person learning continuously.

How could AGI become ASI?#

Recursive self-improvement, where a system improves its own design and so improves faster, is the best-known proposal. It is a hypothesis, not an observed mechanism, and it is not the only route.

Google DeepMind's paper "From AGI to ASI" (published 2026-06-12) lists four pathways:

1. Scaling AGI: more compute, data and capable copies.
2. AI paradigm shifts: new architectures or training methods.
3. Recursive improvement: AI systems improving AI systems.
4. Multi-agent collectives: ASI emerging from large groups of cooperating AI agents.

The paper also discusses frictions and bottlenecks that could slow these pathways, and treats how much they matter as an open question. Our own list of practical limits: compute, energy, training data, the difficulty of verifying that a change is an improvement, and the time real experiments take. The paper argues that progress may keep accelerating, so change could arrive as a series of transformations rather than one step when AGI appears.

We removed the equation this article previously showed for shrinking cycle times. It implied a known rate of improvement that no source supports.

What would an ASI do that humans cannot?#

Everything here is hypothetical. A system that outperformed the best human researchers across scientific fields might design new materials, find mathematics humans have not, or plan logistics at a scale no organisation can. Whether it would formulate new physical paradigms is a possibility, not a forecast. It would still face physical limits: experiments take time, and some questions need data nobody has.

What are the risks of artificial superintelligence?#

The main concern in the safety literature is not hostility but misaligned goals: a highly capable system pursuing an objective whose edge cases harm people. Bostrom's 2014 book Superintelligence: Paths, Dangers, Strategies sets out this argument, including the orthogonality thesis (intelligence and final goals can vary independently) and instrumental convergence (many goals favour acquiring resources and avoiding shutdown). Both are arguments, not experimental results.

For the field's open problems, see our guides to alignment theories and what superintelligence is.