What is AGI?

Three words get used as if they were three points on one line: AI, then AGI, then superintelligence. They are not. One is a category so broad it has stopped carrying information, one has no agreed definition and the disagreement is doing real commercial work, and one is a scenario rather than a measurement. Anybody who tells you how far along that line we are has quietly picked definitions for all three, and the picking is where the argument actually lives.

The short answer#

AI is the field. AGI usually means a system matching or beating human performance across the full range of cognitive work rather than in one domain, and there is no agreed test for it. Superintelligence means performance beyond the best humans at essentially everything, which nobody can measure because the comparison class does not exist yet.

The useful thing to know is that the second definition is contested by people with money riding on the answer, and that current capability is uneven enough that "how close are we" has no single value. A system can sit at graduate level on one task and below a competent adult on the next.

Why the word AI stopped being useful#

AI names a research field, not a technology. Spam filters are AI. Chess engines are AI. So are the large language models that prompted the current wave, and so are the recommendation systems that have been quietly ranking things for twenty years. When a vendor, a minister or a newspaper says AI, the word is doing almost no work, and the sentence usually survives its removal.

In the current period the term that carries meaning is general-purpose AI: a system trained broadly enough to be pointed at work it was not built for. That is the term the International AI Safety Report uses, and the choice is deliberate, because a definition tied to a capability can be checked and a definition tied to a vibe cannot.

AGI: the definitions do not agree#

The clearest evidence that AGI lacks a settled meaning is that a team at Google DeepMind went through the published definitions and concluded a new framework was needed. Their proposal, presented at ICML in 2024, replaces the threshold with six levels of performance crossed with breadth: No AI, Emerging, Competent, Expert, Virtuoso and Superhuman, where Competent means the 50th percentile of skilled adults, Expert the 90th and Virtuoso the 99th.

Two things follow from that structure. The first is that a system can be Superhuman on narrow tasks and Emerging on general ones at the same moment, so a single answer to "have we got there" is a category error. The second is that a capability level says nothing about how a system should be deployed. The authors treat autonomy as a separate axis, and they are right to: knowing what a model can do does not tell you what it should be allowed to do without a person.

Other definitions are in circulation and they do different work. OpenAI's charter language, "highly autonomous systems that outperform humans at most economically valuable work", is narrower than it sounds, because economically valuable work is not the same as cognitive work and the phrase leaves the threshold open. The survey literature avoids the term altogether and asks about High-Level Machine Intelligence instead, defined as unaided machines accomplishing every task better and more cheaply than human workers, setting aside tasks where being human is itself the point, and judged on feasibility rather than on whether anybody adopts it.

Those three are not variants of one idea. They would be reached at different times, by different systems, and two of them are contractual rather than scientific. When somebody says AGI is five years away, the first question is which of these they mean.

What the forecasts say, and what moves them#

The largest survey of its kind put questions to 2,778 researchers who had published in the previous year at NeurIPS, ICML, ICLR, AAAI, IJCAI or JMLR. On timing, the 2023 aggregate forecast gave High-Level Machine Intelligence a 50 per cent chance by 2047, and a 10 per cent chance by 2027.

The number worth holding is not 2047. It is that the same question one year earlier produced 2060. Thirteen years of expected timeline vanished in twelve months, in a population that had moved the same estimate by a single year across the previous six. Whatever that measures, it is not a stable read on the future.

On risk the survey did something more useful than ask once. Different respondents drawn from the same population got differently worded questions. Asked what probability they put on future AI advances causing human extinction or similarly permanent and severe disempowerment, the median was 5 per cent. Asked about human inability to control advanced AI causing the same outcome, the median was 10 per cent. Same population, same fortnight, one changed clause, double the answer. Depending on the wording, between 41.2 and 51.4 per cent gave more than a one in ten chance.

That is the finding to take to an audience, and it cuts in a direction people do not expect. The doubling does not mean the risk is unreal, and it does not mean the experts are unserious. It means the number is partly an artefact of the question, so anyone quoting a single figure without its wording has dropped the part that determined it. The survey authors say as much themselves, noting that their participants are experts in AI rather than trained forecasters, and citing a related study in which changing the framing moved lay estimates of existential risk by nearly six orders of magnitude.

Superintelligence#

Superintelligence describes a system that outperforms the best human at essentially every cognitive task. In the DeepMind framework it is the top performance band applied at full breadth, and no current system is close to it on the general axis.

The reason it dominates public conversation out of proportion to the evidence is that the arguments about it are arguments about consequences rather than about capability, and consequence arguments do not need a measurement to be made. That makes them impossible to settle and easy to publish. It is reasonable to take the scenario seriously and also to notice that no amount of debating it tells you anything about what a model can do this year.

How close are we#

The most institutionally backed answer available comes from the 2026 International AI Safety Report, produced by more than a hundred experts with an advisory panel nominated by over thirty countries. Its description of current capability is jagged: systems solve graduate-level mathematics and science problems, and fail simpler things; reliability falls away across many steps; hallucination persists; performance drops on the physical world, and on unfamiliar languages and cultural contexts.

On agents, the report finds they complete software engineering tasks with limited oversight but cannot yet sustain the long-horizon planning that automating a whole job requires, and concludes that for now they complement people rather than replace them. On loss of control, it reports that expert views vary widely and that current systems show at most early signs of the relevant behaviours.

The report also names the position decision-makers are actually in, calling it an evidence dilemma: capability moves quickly and evidence about new risks arrives slowly, so acting early risks entrenching the wrong intervention and waiting risks leaving people exposed. That is a more honest description of the choice facing a board than any timeline.

Settle the definition before arguing about the date#

• Ask which definition, before arguing about the date. Most disagreements about AGI timelines are disagreements about the finish line that neither party has stated. Naming it usually ends the argument or makes it a real one.
• Treat a single risk number as incomplete without its wording. Five per cent and ten per cent came from the same researchers in the same survey. Quote the question, not just the figure.
• Plan against jagged capability, not against a threshold. The operational question is which specific tasks a system does well in your setting, tested there. A general answer about how advanced AI has become will not tell you.
• Separate what it can do from what it may do alone. The DeepMind framework keeps capability and autonomy on different axes for a reason. Most governance failures collapse them.

The forecast data points both ways#

Nothing here says AGI is impossible, or far away. The forecast data is a poor instrument, and a poor instrument pointing at a long timeline is no more reassuring than one pointing at a short one. It also says nothing about whether the current approach scales to general capability, which is a technical question the evidence does not answer.

The risk literature has a second problem this page does not solve. The people most willing to give a number are the people who have thought hardest about the scenario, which is a selection effect running in the direction of higher estimates, and the people most dismissive rarely give a number at all, so their view never enters the average. Both sides of that are unmeasured.