SiPoly

The AI forecaster scoreboard & promise tracker

AGI vs superintelligence

AGI and superintelligence are two rungs on the same ladder, and confusing them is the most common error in the public debate. The distinction is simple: AGI is an AI as capable as a human across the board; superintelligence is an AI vastly more capable than the best humans at everything. One is a peer. The other is a successor.

That gap — between “as good as us” and “far beyond us” — is where most of the policy argument lives. An AGI era would be disruptive in the way the industrial revolution was disruptive: jobs change, power shifts, new rules are needed. A superintelligence era would be disruptive in a different way entirely: the systems making decisions would be better at the decisions than we are, including decisions about themselves.

Two rungs on the same ladder

Think of it as three levels. Narrow AI is what we have today: systems that are superhuman at one thing (chess, translation, protein folding) and useless at everything else. AGI is the next level: a system that can do, broadly, the intellectual work a human can do — learn new skills, reason across domains, handle open-ended tasks — at roughly human level. Superintelligence is the level after that: an intellect that, in Nick Bostrom’s definition, “greatly exceeds the cognitive performance of humans in virtually all domains of interest.”

The ladder is not a law of nature. It is a way of talking about capability, and different writers draw the rungs in different places. But the ordering matters: almost everyone who thinks superintelligence is possible thinks AGI comes first, because a system has to reach human level before it can pass it. The philosopher David Chalmers breaks the path into equivalence (as good as us), extension (better than us), and amplification (dominating us across arbitrary tasks).

The definitions in use

Because there is no official test for either term, the definitions people actually use are worth knowing:

  • OpenAI’s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” Note the word economically: this is a definition about jobs and value, not about raw intellect.
  • Google DeepMind’s “Levels of AGI” (Morris et al., 2023) proposes a ladder based on two dimensions — depth (how well the system performs) and breadth (how generally). A system could be expert-level but narrow, or competent but broad. AGI, in this framework, is high on both.
  • Anthropic’s Dario Amodei dislikes the term AGI and defines “powerful AI” instead: a system smarter than a Nobel Prize winner across most fields, able to work autonomously on tasks lasting days or weeks, with millions of copies running in parallel. That description is closer to superintelligence than to a bare human-level AGI — which is part of why his timelines sound so aggressive.
  • Bostrom’s superintelligence is defined by a large margin over humans, not a small one: not “somewhat better,” but “greatly exceeds.” A system that is merely 10% better than the best human economist is not superintelligence. A system that is better than every human at everything, and can run a million copies of itself at a hundred times human speed, is.

The practical consequence: when a lab leader says “AGI by 2027,” ask which definition they are using. When someone says “superintelligence,” ask how large the margin is supposed to be.

Why the line is blurry

If the distinction is so simple, why is it so hard to apply? Four reasons:

  1. There is no agreed test. There is no exam you can give a system to certify AGI, let alone superintelligence. Benchmarks measure narrow skills; real generality shows up in handling the unexpected. Any test can be gamed, and any threshold can be argued about.
  2. “Outperform” is doing a lot of work. Outperform at what, measured how, and for whom? A system might outperform humans on a written test while failing at the same task in the real world, where the task involves persuasion, physical action, or dealing with ambiguity.
  3. Capability is not a single number. A system can be superhuman at mathematics and subhuman at common sense. Is that AGI? Superintelligence? Neither? The DeepMind levels framework exists precisely because one number cannot capture this.
  4. The declaration problem. Someone has to decide when the line has been crossed, and the people with the most information — the labs — also have the most to gain from declaring it. Journalist Kevin Roose predicted that an AI company would claim AGI “probably in 2026 or 2027.” Whether a claim or an independent evaluation settles the question matters enormously.

This blurriness is not a pedantic problem. It is the central policy problem: you cannot regulate a threshold you cannot measure.

Why the distinction matters for policy

The two levels call for different policies, on different timelines:

  • In an AGI world, the urgent questions are economic and institutional: what happens to knowledge work, who owns the systems, how do you retrain a workforce and tax the output? These are hard but familiar problems, and the policy toolkit (labour policy, competition law, education) exists.
  • In a superintelligence world, the urgent questions are about control. Can we steer a system smarter than us? Can we verify that it is doing what we asked? Can we keep it from concentrating power irreversibly, or from being copied and misused? OpenAI’s superalignment programme was founded on the premise that our current techniques — which rely on humans supervising AI — will not scale to systems smarter than their supervisors. That is a superintelligence problem, not an AGI problem.

The timeline question sharpens this. If AGI arrives years before superintelligence, society gets a rehearsal: we learn to live with human-level systems, build institutions, and make mistakes at survivable scale. If the gap is months — the fast-takeoff scenario — there is no rehearsal. That is why the takeoff debate, covered in our companion entry, is really a debate about how much time the distinction buys us.

Could we get one without the other?

Mostly, the field assumes AGI first, superintelligence second. But a system could be superhuman in one domain (say, AI research) while remaining subhuman in others — superintelligence in a slice, AGI nowhere. Conversely, a system could reach broad human level and stay there for a long time if progress hits bottlenecks in data, energy, or ideas. The ladder is a useful map, but the territory may not cooperate.

What the scoreboard says

SiPoly’s AGI & superintelligence timelines topic tracks both kinds of claim, and the grading record shows the definitional problem in action. Sam Altman’s claim that superintelligence would arrive “in a few thousand days” was listed as too vague to grade — no definition, no date, nothing to test (claim page). Dario Amodei’s “country of geniuses in a data center,” by contrast, is gradeable precisely because he defined it operationally: smarter than top human experts across most fields, autonomous on multi-day tasks, many parallel copies, by September 2028 (claim page).

On the AGI side, Gary Marcus’s “No AGI in 2025” was graded correct (claim page), and Elon Musk’s “AGI within two years” (April 2024) was graded wrong. The pattern is consistent: AGI claims usually come with dates and get graded; superintelligence claims often fail at the first hurdle, because the speaker never said what would count.

Across the whole scoreboard, as of 5 November 2026, 367 claims are tracked, 129 resolved, and 60% came true.

Key sources

Common misconceptions

  • “AGI and superintelligence are synonyms.” They are not. AGI is the threshold of human-level generality; superintelligence is a large margin beyond it. Conflating them makes every timeline debate incoherent.
  • “AGI will automatically become superintelligence overnight.” Only under a fast takeoff. If progress is limited by experiments, data, energy, and institutions, the AGI era could last years or decades — which is exactly what the takeoff debate is about.
  • “Superintelligence is just AGI with more computing power.” Scale is one route, but self-improvement, speed, and the ability to run millions of copies are equally important. A system does not need a bigger brain to be superintelligent; a human-level brain running a million times faster, in a million copies, would already qualify.

See also