What is AGI?

The word, and where it comes from

“Artificial general intelligence” was coined to draw a line between two kinds of system. On one side is narrow AI: software that does one thing, or a set of related things, very well: recognising faces, translating, playing chess, predicting protein structures. On the other is a system that could turn its hand to anything a person can: learn a new skill from instructions, reason about an unfamiliar problem, plan over days, and move between domains without being rebuilt for each one.

The phrase came into wide use through a 2007 collection edited by Ben Goertzel and Cassio Pennachin, though the idea is older than the label. It was, for most of its life, a term used by a small research community. It is now used by the largest technology companies to describe what they say they are building.

There is no agreed test for AGI. Alan Turing’s 1950 “imitation game” is often cited but is a test of conversation, not general competence, and current chatbots arguably pass it. Companies and researchers use their own working definitions (OpenAI’s charter describes AGI as “highly autonomous systems that outperform humans at most economically valuable work”; others prefer “can do any cognitive task a human can”), and these definitions do not agree with each other. That is the main reason people can look at the same system and disagree about whether it is close.

How AGI relates to the other two terms

AI

Software that does tasks we associate with intelligence. It exists, and has for decades.

AGI

Roughly any intellectual task a person can do. No current system is universally accepted as this; whether today's systems are close is disputed.

Superintelligence

Greatly exceeds the best human minds in virtually all domains. Hypothetical; no such system exists.

The distinction between the second and third rungs matters. A system that matched a competent human across the board would be AGI. A system that made the best human in every field look like a novice would be superintelligence, in the sense the philosopher Nick Bostrom gave the word in 2014. Much of the public argument about AI risk concerns the third rung, and whether reaching the second leads quickly to it, though safety work also covers misuse, accidents and other harms from systems well short of either.

What AGI is not: the administration’s “Super Intelligence”

Since September 2026 there has been a fourth term in circulation, and it is the one most likely to confuse a reader who has arrived here from the news.

On 19 September 2026 President Trump posted a poll on Truth Social (a social-media poll, not a policy) proposing that “artificial intelligence” be renamed “Superior”, “Extreme” or “Supreme” Intelligence, and calling the existing term “inaccurate, and very ineloquent”. A second Truth Social poll on 21 September 2026 offered “Super” against “Superior”. On 22 September 2026, in his address to the UN General Assembly, he said: “From this point forward, all of United States documents and hopefully the world’s, will be changed to use the much more accurate term Super as opposed to Artificial.” He added: “Let’s see if I have any power. Maybe I do and maybe I don’t. We’re gonna find out pretty soon.” The White House’s own press release on the speech, published on 22 September 2026, refers to “the Artificial Intelligence being spoken of so much now — hereinafter officially called ‘Super Intelligence.’”

Three things follow for anyone trying to understand AGI.

First, “Super Intelligence” in this usage is a replacement name for AI as a whole. It covers the chatbot on your phone and the spam filter in your email. It does not describe a new capability, and nothing published by the administration claims that AGI or superintelligence has been reached.

Second, the change so far consists of two social-media polls, a statement in a speech, and a White House press release that uses the words “officially called”. Those are three different things, and none of them is an executive order. As of 23 September 2026, we found no executive order, directive or federal style guide implementing the change among the presidential actions published on whitehouse.gov, and no evidence of the term in use across federal documents; Axios reported on 22 September 2026 that the White House had not responded to a request for comment on whether the President intended the existing research meaning of “superintelligence”. This is a finding that can change quickly; it is re-checked on each review of this page, and the rename timeline records each step as it happens: what was polled, what was said, what the White House has published, and what, if anything, became an executive order or actual usage across federal documents.

Third, as of 23 September 2026 the research community had not adopted the term. The companies building the most capable systems continue to describe their goal as AGI, and a small number use the word superintelligence for what they are aiming at, in something like Bostrom’s sense, not the administration’s.

So a reader who first met the phrase “super intelligence” through President Trump’s speech of 22 September 2026 should leave with the ladder above: what exists today is AI, whatever it is called on a government form; AGI is the disputed next rung, with no current system universally accepted as having reached it; superintelligence is the hypothetical one above that. The administration’s terminology does not demonstrate that either has been achieved.

Where the argument actually is

Whether AGI is near is a live dispute among people who build these systems, and it is worth knowing the shape of it rather than any single forecast.

The case that it is close rests on the pace of the last four years: systems that in 2022 could not reliably do arithmetic now pass professional examinations, write working software and carry out multi-step tasks with tools. Some company leaders have put dates on it. The case that it is not close rests on the things current systems still cannot do (learn continuously from experience, act reliably over long horizons, know when they are wrong) and on the argument that the current method may not get there without a further breakthrough. The International AI Safety Report 2026, published on 3 February 2026 and the nearest thing to a consensus account, says that “general-purpose AI capabilities have continued to improve”, that “experts disagree about whether these capabilities will generalise beyond mathematics and programming”, and that by 2030 “AI progress could plausibly range from stagnation to rapid improvement to levels that exceed human cognitive performance”.

When could superintelligence happen? sets out the forecasts. AGI vs superintelligence: what’s the difference? goes further into the two upper rungs.

How this site uses the words

• AI for the technology in general use.
• AGI for the hypothetical human-level-general system.
• Superintelligence only in the research sense.
• “Super Intelligence”, in quotation marks, when reporting the US administration’s usage.