Outperforming humans at most jobs.
OpenAI’s charter (2018) aims at “highly autonomous systems” that outperform humans at most economically valuable work. The definition is explicitly economic: it is measured by the substitution of human work.
Understanding AI at work / Short page
Understand · Short page
AGI: what are we talking about?
Several laboratories say they are trying to build artificial general intelligence. This page summarises the competing definitions, what the measurements show in November 2026 and what that implies for work.
The essentials
A stated goal, a disputed notion.
What can be said without betting on a date.
Direct answer
Artificial general intelligence, or AGI, is a goal stated by several laboratories. There is neither a shared definition nor a test that would establish its arrival.
Announcements are not measurements: measured, dated indicators are more informative than statements from players engaged in a race for funding.
01 / Definitions
Three ways of defining it.
Each sheds light on what is at stake for work.
A progression rather than a threshold.
Google DeepMind crosses the level of performance, from “emerging” to “superhuman”, with the breadth of tasks covered. The degree of autonomy left to the system is treated separately: it is a deployment choice, not only a capability.
Comparing with the abilities of an educated adult.
A group of researchers measures models on ten broad abilities of human cognition. GPT-4 scored 27% and GPT-5 57%, with a very uneven profile: extensive knowledge, almost no long-term memory.
02 / Dated reference points
What the measurements say, and the announcements.
Four reference points, to read with their source.
| Reference | What it indicates | Source |
|---|---|---|
| Cognitive profile | Across ten abilities each scored out of 10, GPT-5 reaches 57%. Long-term memory storage remains at zero: the model learns nothing lasting from one conversation to the next, unless the tool adds an external memory. As of 13 November 2026, no more recent model has been assessed with this grid. | Hendrycks, Bengio et al., 2025 ↗ |
| Levels of AGI | In 2023, the authors classed general AI at the “emerging” level; the “competent” level had not been reached. More recent models have not been officially reclassified. | Morris et al., 2023 ↗ |
| Announcements | On 3 September 2026, OpenAI’s president said he personally considered AGI to have been reached; nine days later, the head of Anthropic called for the pace of progress to slow. | Axios, 2026 ↗ · Dario Amodei, 2026 ↗ |
| Forecasts | The International AI Safety Report (February 2026) considers both a plateau and a sharp acceleration plausible by 2030. | International AI Safety Report 2026 ↗ |
03 / Debate
Two readings of the same curves.
The arguments on each side, without settling the matter.
Why some expect an acceleration.
Measured capabilities have progressed steadily for several years. Tests designed to resist models fall quickly: ARC-AGI-3 went from under 1% to 81% in less than eight months. The compute and investment devoted to training are still rising sharply.
Why others remain sceptical.
Continual learning and lasting memory remain very limited. Scores depend on the protocol. Reliability on long tasks is progressing more slowly than the demonstrations. Passing tests does not guarantee holding a job in a real environment.
04 / At work
For occupational health, the issue is not the date.
What is worth tracking from now on.
The issue is not to bet on a date. It is to track, task by task, what is becoming automatable, at what pace and under what control, and to anticipate the effects on workload, autonomy, skills and employment before they impose themselves.
OpenAI’s definition is measured by the substitution of human work: that is the scenario examined in the pathway AI and jobs.
Key answers
Three frequent questions.
Short answers, drawn from the guide.
Does AGI already exist?
Nothing allows that to be established: there is neither a shared definition nor an arrival test. OpenAI’s president said in September 2026 that he personally considered it reached. On the cognitive grid of Hendrycks, Bengio and colleagues, the last model assessed, GPT-5, scored 57%.
When will AGI arrive?
Nobody knows. Several laboratory leaders publicly mention a horizon of a few years; the International AI Safety Report considers both a plateau and a sharp acceleration plausible by 2030.
What is the difference between AGI and superintelligence?
In Google DeepMind’s grid, superintelligence is the last level of a general AI: doing better than all humans on many tasks. In 2023, the authors classed it as not reached.
Where this page comes from
It summarises one chapter of the guide Understanding AI at work, which gives the source, date and limits of every figure. The figures dated November 2026 will go out of date: the guide is the version kept current.
Update log
- First version, drawn from the guide “Understanding AI at work” updated on 13 November 2026.