AGI vs ASI: Human-Level vs Superhuman AI
Where artificial general intelligence ends and artificial superintelligence begins on the Levels of AGI scale.
AGI (artificial general intelligence) would match human ability across most tasks, and ASI (artificial superintelligence) would exceed the best humans at nearly all of them.
At Layer3Labs, we build AI workflows for small and mid-sized businesses using today's AI, which sits below both thresholds.
Neither system exists as of October 2026. According to IBM, current LLMs are not considered AGI, though some researchers argue advanced LLMs already qualify.
AGI (artificial general intelligence) vs. ASI (artificial superintelligence): Side-by-Side
| Dimension | AGI (artificial general intelligence) | ASI (artificial superintelligence) |
|---|---|---|
| Definition | A hypothetical stage where an autonomous system matches or exceeds human cognitive abilities across any task or economically valuable work. | A hypothetical software-based system with an intellect that greatly exceeds the cognitive performance of humans across virtually all domains. |
| Capability bar versus humans | Matches the cognitive abilities of skilled human adults across general domains or outperforms humans at most economically valuable work. | Outperforms 100% of humans across all domains, exceeding human intelligence in virtually every field of interest. |
| Position on the Levels of AGI scale | Spans Level 1 Emerging to Level 4 Exceptional general AI in the Morris et al. framework, with Level 2 Competent best corresponding to historical conceptions. | Occupies Level 5 Superhuman general AI in the Morris et al. framework, outperforming all humans across all tasks. |
| Exists today? | No agreed test or consensus says any system has achieved it; the Levels of AGI paper says no public system had reached Competent AGI at the time of writing. | No system exists today; researchers and industry documentation describe it as a hypothetical future state. |
| Where today's chatbots sit | The 2023 Morris et al. paper classified ChatGPT, Bard, Llama 2 and Gemini as Level 1 Emerging AGI; it does not rate newer models such as GPT-6 Astra. | Chatbots do not qualify as superintelligence, which requires outperforming all humans across all general cognitive domains. |
| Who is pursuing it | OpenAI, whose charter defines AGI; its agreement with Microsoft adds an independent expert panel to verify any AGI declaration by OpenAI. | OpenAI's leadership wrote in 2023 about governing superintelligence; Meta Superintelligence Labs and Safe Superintelligence Inc. use the word in their names. |
| Common definitions in use | The OpenAI Charter defines it as outperforming humans at most economically valuable work, while IBM defines it as matching or exceeding cognitive abilities across any task. | Nick Bostrom defines it as an intellect greatly exceeding humans in virtually all domains, while IBM defines it as software with an intellectual scope beyond human intelligence. |
| Main concern raised | Verification disputes and the absence of an agreed test or consensus that it exists. | Whether it can be built safely and controllably; the Statement on Superintelligence, with more than 77,000 signatures as of October 2026, calls for a prohibition until there is broad scientific consensus on that and strong public buy-in. |
Are you one of these vendors? Update your listing
AGI vs ASI: The Core Difference
The core difference between artificial general intelligence (AGI) and artificial superintelligence (ASI) centers on whether machine intellect matches human capability or surpasses every human mind.
Both terms describe hypothetical stages of technological development rather than existing commercial software. According to research documentation from IBM on AGI, AGI represents a hypothetical milestone in machine learning where software matches or exceeds human cognitive abilities across any task. In contrast, IBM's analysis of ASI characterizes superintelligence as a hypothetical software-based system with "an intellectual scope beyond human intelligence."
Outside computer science discussions, the abbreviation AGI carries an unrelated legal meaning. In United States tax law, AGI stands for adjusted gross income, which the Internal Revenue Service (IRS) defines as "your total (gross) income from all sources" minus certain adjustments.
- Artificial general intelligence (AGI) targets parity with skilled human intellect across broad general work.
- Artificial superintelligence (ASI) describes machine cognitive performance that surpasses all human capabilities across every domain.
- Neither AGI nor ASI exists as of October 2026.
Three Developmental Stages from Narrow AI to Superintelligence
A common shorthand splits AI into three stages: narrow AI (ANI), AGI and ASI. The Levels of AGI paper gives that shorthand a scale. Its narrow track runs from calculators to AlphaFold, and its AGI levels start at Emerging AGI, where it placed ChatGPT at the time of writing. General AI that outperforms all humans is ASI.
The second stage, artificial general intelligence, represents general rather than narrow AI capability. An AGI system would match or exceed human cognitive abilities across any task and outperform humans at most economically valuable work. Our AI vs AGI comparison and what is AGI guide go further.
The third stage, artificial superintelligence, represents an intellect that dramatically outstrips humanity in every intellectual endeavor. OpenAI's 2023 post describes superintelligence as "dramatically more capable than even AGI," and the Levels of AGI paper puts ASI one level above Exceptional AGI, so both place ASI after AGI. Our what is superintelligence guide covers ASI on its own, and AI vs SI explains why the federal SI label is a different thing.
Capability Thresholds on the Google DeepMind Scale
The Google DeepMind Levels of AGI framework formalizes system progression by combining performance depth with general versus narrow scope. Morris et al., with Shane Legg as a co-author, posted the paper on arXiv in November 2023 and revised it in September 2025. Its matrix defines six performance tiers, from Level 0 to Level 5. The framework emphasizes that intelligence requires evaluating both task breadth and raw performance percentiles relative to human adults.
Level 0 represents No AI, encompassing calculators in narrow applications and crowdsourced computing like Mechanical Turk in general domains. Level 1 represents Emerging systems that perform equal to or somewhat better than an unskilled human adult. The paper lists good old-fashioned AI (GOFAI) and SHRDLU as narrow Level 1 systems. Its general Level 1 examples at the time of writing were ChatGPT, Bard, Llama 2 and Gemini.
The higher tiers establish strict criteria for true general capability. Level 2 Competent requires matching at least the 50th percentile of skilled adults, Level 3 Expert requires matching the 90th percentile, and Level 4 Exceptional requires reaching the 99th percentile. Morris et al. state that Competent AGI "has not been achieved by any public systems at the time of writing" and "best corresponds to many prior conceptions of AGI." Level 5 Superhuman means outperforming 100% of humans. AlphaFold, AlphaZero and Stockfish are the paper's narrow Level 5 examples, and general Level 5 is artificial superintelligence, which no system has reached.
Technical Definitions Across Industry Charters and Benchmarks
Industry organizations, academic literature, and evaluation benchmarks define general intelligence using distinct operational criteria. Wikipedia records that Mark Gubrud used the term artificial general intelligence in 1997 and that Shane Legg and Ben Goertzel popularized it in the 2000s. IBM credits Goertzel with popularizing it in 2007 at Legg's suggestion.
The published OpenAI Charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work. Under a formal partnership agreement announced on Microsoft's official blog in October 2025, once AGI is declared by OpenAI, that declaration will now be verified by an independent expert panel.
The ARC Prize ARC-AGI benchmark, developed from François Chollet's 2019 research, measures skill-acquisition efficiency on unfamiliar tasks. It tests fluid intelligence rather than stored knowledge, across ARC-AGI-1, ARC-AGI-2 and ARC-AGI-3.
Where ChatGPT and Other Chatbots Sit in the AGI Debate
The 2023 Levels of AGI paper rated the chatbots it reviewed (ChatGPT, Bard, Llama 2 and Gemini) as Level 1 Emerging AGI, the lowest level of its general track. It said no public system had reached Level 2 Competent AGI at the time of writing. According to IBM's research summary, current large language models are not considered AGI, although some researchers argue advanced models already qualify while many others disagree.
Wikipedia calls whether AGI exists "a subject of intense debate within the AI community" and records no agreed test or consensus that any system has achieved it.
The Levels of AGI paper gives no rating for models released since; OpenAI's claims about GPT-6 Astra are covered in Is GPT-6 Astra AGI?.
The Intelligence Explosion and Governance Concerns Surrounding Superintelligence
In 1965, I.J. Good argued that "an ultraintelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion'" (Wikipedia).
In a May 2023 post, Governance of superintelligence, Sam Altman, Greg Brockman, and Ilya Sutskever described superintelligence as future AI systems "dramatically more capable than even AGI." They wrote that "it's conceivable that within the next ten years, AI systems will exceed expert skill level in most domains."
On October 20, 2025, the Future of Life Institute published the Statement on Superintelligence. It calls for "a prohibition on the development of superintelligence, not lifted before there is broad scientific consensus that it will be done safely and controllably, and strong public buy-in." The statement had more than 77,000 signatures as of October 2026.
Federal Policy Terminology and the Executive Order Definition
Executive Order 14434, signed September 29, 2026, tells executive-branch agencies to say "Super Intelligence" (SI) instead of "AI" in their own non-statutory documents. It defines SI by the statutory definition of artificial intelligence at 15 U.S.C. 9401(3). Because SI carries that statutory AI definition, federal SI means today's AI. The order never mentions AGI or ASI; AI vs SI covers the difference.
Evaluation Criteria for Future Capability Claims
Test any AGI claim with two questions. When a lab announces an AGI milestone, ask whether it was tested on unfamiliar tasks, as ARC-AGI does, and whether anyone outside the lab verified it.
Our view that neither AGI nor ASI exists today would change if the independent expert panel in the Microsoft-OpenAI agreement verified an AGI declaration. It would also change if a public system sustained Level 2 Competent performance across many kinds of tasks.
IBM says current LLMs are not considered AGI and calls ASI "still a hypothetical future state." If you are planning AI work now, test today's models on your own tasks instead of waiting for either milestone.
The Verdict
Artificial general intelligence represents machine capability that matches skilled human adults across broad, economically valuable disciplines. Artificial superintelligence represents a theoretical condition where machine intellect outperforms all human minds across every intellectual and scientific domain.
Neither technology exists as of October 2026. The 2023 Levels of AGI paper rated ChatGPT and Gemini at Level 1 Emerging AGI, equal to or somewhat better than an unskilled human. It said no public system had reached Level 2 Competent AGI at the time of writing, and it puts superintelligence at general Level 5.
For technical leaders and operators, immediate planning should focus on narrow AI capabilities rather than speculative future milestones. Watch ARC-AGI results and any finding from the Microsoft-OpenAI expert panel to track progress toward AGI. Treat the SI label in Executive Order 14434 as the federal name for today's AI; it says nothing about superintelligence.
Researched from primary vendor documentation and public regulator sources. Pricing and availability are accurate as of Oct 9, 2026 and can change — confirm current terms with each vendor before you buy.
Frequently Asked Questions
- AGI refers to artificial general intelligence, a hypothetical AI system that matches or exceeds human cognitive abilities across any task or economically valuable work. ASI refers to artificial superintelligence, a hypothetical future system possessing an intellect that greatly exceeds the cognitive performance of all humans across virtually all domains.
- A common shorthand names three stages: artificial narrow intelligence (ANI), artificial general intelligence (AGI) and artificial superintelligence (ASI). Narrow AI handles specific tasks, AGI would match human ability across most tasks, and ASI would exceed all humans in virtually all domains. The Levels of AGI paper maps these to its narrow track, general Levels 1 to 4, and general Level 5.
- ChatGPT is AI, and IBM says current LLMs are not considered AGI, though some researchers argue advanced LLMs already qualify. The 2023 Levels of AGI paper rated ChatGPT Level 1 Emerging AGI, equal to or somewhat better than an unskilled human. It said no public system had reached Competent AGI at the time of writing.
- No, superintelligence is not the same as AGI. AGI matches skilled human adults across broad cognitive work, whereas superintelligence (ASI) refers to future systems that dramatically exceed the intellectual capabilities of all humans across virtually all fields of interest.
- AGI comes first. OpenAI's 2023 post describes superintelligence as "dramatically more capable than even AGI," and the Levels of AGI paper puts ASI one level above Exceptional AGI. Neither source gives a date for either one.
Need Help Implementing Production-Ready AI Workflows Today?
We design, integrate, and run AI automations across CRMs, document pipelines, and core operations. Book a consultation to map your operational stack.
Book a Consultation