New · France 2027 presidential election: what the candidates propose on AI, quoted and sourced. Explore the tracker →

Last reviewed:

What is AGI? Definition of artificial general intelligence and difference with narrow AI

AGI (artificial general intelligence) refers to an AI system able to match or exceed humans at most cognitive tasks, not just one. There is neither a shared definition nor an accepted test to say it has been reached: each lab uses its own.

The term contrasts with so-called weak, or narrow, AI, specialised in one task (recognising an image, playing chess, sorting emails). The weak versus strong AI distinction comes from philosopher John Searle (1980), for whom a strong AI would really have a mind; AGI, by contrast, is defined by capabilities, with no claim about consciousness. Definitions differ between players. OpenAI's Charter (2018) speaks of “highly autonomous systems that outperform humans at most economically valuable work”. Google DeepMind uses an AI “at least as capable as humans at most cognitive tasks” and proposes a scale of levels crossing performance and generality (Morris et al., 2023). Dario Amodei, CEO of Anthropic, wrote in October 2024 that he dislikes the term and speaks of “powerful AI”: smarter than a Nobel Prize winner across most fields. On timing, lab leaders point to short horizons: DeepMind wrote in April 2025 that AGI “could be here within the coming years”, and Dario Amodei mentioned 2026 at the earliest. Academic research is more cautious: in the AAAI survey published in March 2025 (475 respondents), 76% consider it “unlikely” or “very unlikely” that scaling up current approaches will be enough to reach AGI. The stakes are also contractual: the Microsoft-OpenAI agreement long made commercial rights depend on declaring AGI, before the clause was neutralised on 27 April 2026.

Concrete example

Illustrative case: in 2026, the board of a French mid-sized industrial company asks whether it should “wait for AGI” before investing in AI. The CEO presents three points. First, the tools available today (assistants, agents, reasoning models) are not AGI under any public definition, yet they already automate a measurable share of the company's administrative work (quotes, reminders, summaries). Second, dates announced by labs are not commitments and researchers are divided. Third, reversible choices (short contracts, structured data, in-house skills) remain useful whatever the scenario. The board approves an 18-month plan based on quantified use cases, with no assumption about AGI.

Comparison

Narrow AI, general AI and superintelligence: key points
Narrow AI (weak)General AI (AGI)Superintelligence (ASI)
DefinitionSpecialised in one task or domainMatches or exceeds humans at most cognitive tasksGreatly exceeds humans in virtually all domains
Status at 30/09/2026In service everywhereNot reached under any recognised definition; no shared testHypothetical
ExamplesSpam filter, image recognition, current AI assistantsNo proven exampleNone (fiction only)
Who talks about itThe whole industryLabs (OpenAI, Google DeepMind, Anthropic), researchersPhilosophers (Bostrom), labs in their safety texts
For the companyMeasurable use cases todaySomething to monitor, not a planning assumptionPublic debate topic, no operational effect

FAQ

What is artificial general intelligence?

It is an AI system that would be able to match or exceed humans at most cognitive tasks, not only at one specialised task. No definition is shared by all: OpenAI, Google DeepMind and Anthropic each use their own.

What is the difference between weak AI and strong AI?

Weak, or narrow, AI is specialised in one task: it is all the AI in service today. Strong AI, a notion from philosopher John Searle (1980), refers to a machine that would really have a mind. AGI is defined by the breadth of capabilities, without settling the question of consciousness.

Does AGI already exist?

As of 30 September 2026, no lab or institution has established that a system meets a recognised definition of AGI. Current models perform very well on many tasks but remain uneven, and there is no test accepted by all.

What is an example of AGI?

There is no proven example. ChatGPT, Claude or Gemini are general-purpose models, but they do not meet any public definition of AGI. Frequently cited examples (HAL 9000, Samantha in “Her”) are fiction.

What is the difference between AGI and superintelligence?

AGI refers to a level comparable to humans at most tasks. Superintelligence (ASI), in the definition popularised by philosopher Nick Bostrom (2014), refers to an intellect that would greatly exceed humans in virtually all domains. Both remain hypothetical.

When will AGI arrive?

Nobody knows. Some lab leaders point to a few years (Google DeepMind in April 2025, Dario Amodei for whom 2026 was possible at the earliest). In the March 2025 AAAI survey, 76% of researchers consider it unlikely that scaling up current methods will be enough.

See also

Further reading

Levels of AGI for Operationalizing Progress on the Path to AGI, Morris et al., Google DeepMind, arXiv:2311.02462 (external resource)

Sources

  1. OpenAI Charter, OpenAI, April 2018 (definition of AGI). https://openai.com/charter/ (accessed 2026-09-30)
  2. Taking a responsible path to AGI, Google DeepMind, 2 April 2025. https://deepmind.google/discover/blog/taking-a-responsible-path-to-agi/ (accessed 2026-09-30)
  3. Machines of Loving Grace, Dario Amodei, October 2024. https://www.darioamodei.com/essay/machines-of-loving-grace (accessed 2026-09-30)
  4. AAAI 2025 Presidential Panel on the Future of AI Research, AAAI, March 2025 (survey of 475 respondents). https://aaai.org/wp-content/uploads/2025/03/AAAI-2025-PresPanel-Report-FINAL.pdf (accessed 2026-09-30)

← Back to glossary

Address copied