Artificial general intelligence (AGI) has no single accepted definition, and the term covers at least nine competing ideas, from Alan Turing's 1950 imitation test to a 2018 charter phrase about outperforming humans at most economically valuable work. A 2023 research paper proposed one way to compare them: rate a system by how well it performs against skilled adults and by how many kinds of tasks it handles. On that scale its authors placed the leading chatbots of September 2023 at Level 1, "Emerging AGI."
Key takeaways
- "AGI" is used for several different claims: passing a test, matching a brain, doing most human cognitive tasks, or earning money.
- The "Levels of AGI" paper (first posted November 4, 2023) rates systems on performance and generality, with thresholds at the 50th, 90th, and 99th percentiles of skilled adults.
- The paper judges capability, not deployment, so a system can count as capable of a task even if no one uses it for that task.
- The authors state that their placement of current systems is approximate and that a standardized benchmark for generality does not yet exist.
Why the label is slippery
Computer scientists use "narrow" to describe software built for one task and "general" for software that can handle many. The trouble starts when someone says a system has reached AGI, because the speaker may mean any of several things. A claim about passing a conversation test, a claim about matching human brain complexity, and a claim about replacing paid labor can each be true or false independently of the others. Without a stated definition, two people can disagree about a news story while using different standards.
The paper "Levels of AGI for Operationalizing Progress on the Path to AGI" (Meredith Ringel Morris, Jascha Sohl-Dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra Faust, Clement Farabet, and Shane Legg) reviews nine earlier definitions before offering its own. Its fifth version is dated September 24, 2025, and the placements discussed below come from that text.
Nine definitions the paper reviews
The table paraphrases the paper's case studies and the objection it raises to each.
| Idea (source, year) | What it says in short | Problem the authors raise |
|---|---|---|
| Turing Test (Alan Turing, 1950) | A judge cannot tell machine text from human text | Shows how easily people are fooled, and modern language models pass some versions |
| Strong AI (John Searle, 1980) | A computer running the right software would have a mind | No scientific consensus on how to test for consciousness |
| Brain analogy (Mark Gubrud, 1997) | Systems rival the brain in complexity and speed | Brain-like processes are not strictly necessary |
| Human-level cognitive tasks (Shane Legg and Ben Goertzel, about 2001) | A machine can do the cognitive tasks people usually do | Unclear which tasks and which people |
| Ability to learn tasks (Murray Shanahan, 2015) | Not specialized, and able to learn as wide a range of tasks as a human | Makes learning itself a required skill |
| OpenAI charter (2018) | Highly autonomous systems that outperform humans at most economically valuable work | Economic value can miss skills such as creativity or emotional intelligence |
| Flexible, general intelligence (Gary Marcus, 2022) | Human-comparable reliability on open tasks, such as cooking in an unfamiliar kitchen | Unclear whether the tasks are necessary or sufficient, and some require a body |
| Artificial Capable Intelligence (Mustafa Suleyman, 2023) | Completes complex, multi-step real-world tasks, tested by turning USD 100,000 into USD 1 million | Realistic but narrow, and a pure profit focus could raise alignment risks |
| Language models as generalists (Agüera y Arcas and Norvig, 2023) | Current language models already count, because generality is the key property | Generality needs to be paired with performance |
The levels scale, step by step
The authors derive six principles from this review. Among them: judge capabilities rather than internal processes, weigh both breadth and depth, focus on cognitive and metacognitive tasks (physical ability adds generality but is not required), and measure potential rather than deployment. Deployment is excluded because legal, ethical, or social hurdles can keep a capable system from being used.
Performance is rated against skilled adults, with generality as a second axis.
| Level | Performance threshold | Narrow example named in the paper |
|---|---|---|
| 0, No AI | Not applicable | Calculators, compilers |
| 1, Emerging | Equal to or somewhat better than an unskilled human | SHRDLU, a rule-based system |
| 2, Competent | At least the 50th percentile of skilled adults | Smart speakers |
| 3, Expert | At least the 90th percentile | Grammar checkers |
| 4, Exceptional | At least the 99th percentile | Deep Blue, AlphaGo |
| 5, Superhuman | Outperforms all humans | AlphaFold, Stockfish |
For general systems, the paper states that leading language models of September 2023 reached Competent performance on some tasks, such as short essays and simple coding, but stayed at Emerging performance on most. It therefore classes them as Level 1 general AI and lists no general system at Level 2 or above.
What the scale does not settle
The paper proposes no benchmark. It says that deciding which tasks define generality, and what share a system must pass, is an open research question, and that unambiguous classification will need a standardized task set. A separate scale for autonomy runs from a human doing everything, through AI as a tool, consultant, collaborator, and expert, to a fully autonomous agent. The authors treat autonomy as its own axis because it interacts with risk and with how a system is deployed.
For readers following headlines, the practical use is a set of questions. Which definition is in play? Against which people is performance measured? Which tasks count? Was the result a demonstration of capability or a report of actual use? A claim that answers none of these cannot be compared with another claim.
FAQ
Is AGI the same as superintelligence?
No. In the levels paper, general AI runs from Emerging to Exceptional, while Level 5, Superhuman, outperforms all humans and corresponds to what the paper calls artificial superintelligence.
Did the paper say ChatGPT is AGI?
It classified ChatGPT, Bard, Llama 2, and Gemini as Emerging AGI, the lowest general level, and said Competent AGI had not been achieved.
Does AGI require a robot body?
Not under this framework. It treats physical ability as something that can add generality without being required, while the cooking test cited in the paper does need a body.
Is there an agreed test for AGI?
No. The authors say a standardized benchmark covering a broad task set would be needed, and they do not supply one.
Related reading
- Whether AI could surpass human intelligence
- Where human and machine intelligence differ
- The debate over AI replacing human intelligence
- AI and human reasoning
More in AI Basics.
Last verified: October 2026
Sources
- Morris et al., Levels of AGI for Operationalizing Progress on the Path to AGI, arXiv 2311.02462, first posted November 4, 2023, version 5 dated September 24, 2025