What is AGI? The letters stand for artificial general intelligence, and the honest first answer is that nobody has agreed on a single definition. The term usually points to a machine that can learn and perform the full range of intellectual tasks a human can, rather than excelling at one narrow job such as chess or translation. Beyond that, the word gets stretched to fit whatever argument the speaker is making. This article sets out the main definitions, what experts and forecasters say about timing, what skeptics argue, and which claims are evidence and which are marketing or guesswork.

What does AGI actually mean?

The phrase has a short history. Physicist Mark Gubrud used the term in 1997, and researchers including Shane Legg and Ben Goertzel popularized it around 2002 to separate the long-range goal of human-level machine intelligence from the narrow systems then called AI. Marcus Hutter’s AIXI formalism in 2000 gave one mathematical idealization of a general agent.

Today the commonly stated meaning is a system that matches or exceeds human capability across most cognitive tasks, including ones it was never specifically trained for. A related, bigger idea is artificial superintelligence, a system that greatly exceeds humans in essentially every domain. Philosopher John Searle used “strong AI” in 1980 to mean something different again: a machine that has a mind and consciousness. Most current debate about AGI sidesteps consciousness and focuses on capability.

That already gives three separate questions. Can a system do what humans do? Can it do it across domains without retraining? And does it experience anything? Evidence on the first is accumulating, the second is disputed, and the third is largely philosophical.

Why can’t experts agree on a definition?

Because the definition quietly decides the answer. Define AGI as “can pass a conversation test” and you may already be close. Define it as “can do every economically valuable human task, including physical ones” and you are much further away.

Google DeepMind researchers proposed a framework in 2023 that treats generality and performance as separate axes, with levels labeled emerging, competent, expert, virtuoso and superhuman. By that scale, they placed the chatbots of the time, including ChatGPT, at the “emerging” level of general capability. The point of the framework is that “AGI or not” is the wrong yes-or-no question. A system can be superhuman at some tasks and unreliable at others.

Other proposed tests vary wildly. Alan Turing’s 1950 imitation game asks whether a judge can tell machine from human in conversation. A 2025 study by Jones and Bergen reported that GPT-4.5 was judged to be the human in 73 percent of five-minute text chats, higher than the 67 percent rate for the real humans. Many researchers regard the Turing test as a measure of conversational mimicry, not general intelligence. Apple co-founder Steve Wozniak suggested a coffee test, in which a machine must enter an ordinary home and figure out how to make coffee. Mustafa Suleyman suggested giving an AI $100,000 and seeing whether it can turn it into $1 million. Each test captures a different slice of “general.”

What have people predicted, and how have those predictions fared?

The early record is humbling. In 1965, Herbert Simon predicted that within twenty years machines would be capable of doing any work a man can do. In 1967, Marvin Minsky said the problem of creating AI would be substantially solved within a generation. Neither happened.

A 2012 analysis by Armstrong and Sotala looked at 95 predictions made between 1950 and 2012 and found a strong tendency to place AGI roughly 15 to 25 years in the future, regardless of when the prediction was made. A separate meta-analysis found a similar pattern, though it has been criticized. The lesson drawn by some is that “about twenty years away” is a recurring habit, not a measurement.

Recent forecasts have moved sharply earlier. A review from the group 80,000 Hours collects several, with caveats about each:

  • A large survey of AI researchers (Grace and colleagues, conducted in 2023 and published in 2024) defined “high-level machine intelligence” as AI doing every task better and more cheaply than humans. The median respondent gave a 25 percent chance by the early 2030s and a 50 percent chance by 2047, 13 years earlier than the same survey’s 2022 edition.
  • The forecasting platform Metaculus, with nearly 2,000 forecasters as of February 2026, put the chance of AGI under its own four-part definition at 25 percent by 2029 and 50 percent by 2033. In 2020 the median was about 50 years away.
  • Leaders of AI companies in early 2025 mostly expected AGI within roughly two to five years.

Each group has weaknesses. Company leaders have a financial interest in showing progress. Academic researchers are not forecasting specialists and gave inconsistent answers to some questions in the survey. Forecasting communities are self-selected for interest in AI. The review’s author concluded that expert opinion neither rules AGI in nor out before 2030, and that no group is clearly reliable.

What do prominent figures claim?

Claims range widely, and dates matter. Geoffrey Hinton said he once thought AGI was 30 to 50 years or more away, changed his mind in 2023, and in 2024 estimated, with low confidence, five to twenty years. Demis Hassabis said in May 2023 he expected AGI within a decade, possibly in a few years. Former OpenAI researcher Leopold Aschenbrenner wrote in June 2024 that AGI by 2027 was “strikingly plausible.”

Some go further and say it has partly arrived. In 2023 a group of Microsoft researchers called GPT-4 an early but incomplete form of AGI, and the same year Blaise Agüera y Arcas and Peter Norvig argued that frontier models already show significant general intelligence. In December 2025, OpenAI chief executive Sam Altman said that AGI “kinda went whooshing by” with less impact than people expected, and suggested shifting attention to superintelligence.

That last statement shows the definitional problem in action. If the word can be declared achieved with little visible change in the world, it may say more about where the goalposts sit than about the technology.

What do skeptics say?

Skeptics come in several types. Some argue the goal is simply far away. Paul Allen wrote in 2011 that AGI was unlikely this century because it would require unforeseeable breakthroughs. Alan Winfield compared the gap to that between current spaceflight and faster-than-light travel.

Others object to the method. Yann LeCun, a pioneer of deep learning, said in 2025 that large language models are a dead end for reaching superintelligence, and he left Meta later that year to start a company, announced as Advanced Machine Intelligence Labs, building “world models” that learn the structure of the physical world rather than only predicting text. Some embodied-cognition critics argue that human intelligence depends on having a body acting in an environment, so a disembodied model may be missing something essential. Philosophers Hubert Dreyfus and Roger Penrose argued on different grounds that strong AI is not achievable.

A third group questions the motives behind the rhetoric. Andrew Ng has suggested that some warnings about existential risk from AI companies may be hype or an attempt to shape regulation in their favor. Skeptics also argue that talk of future superintelligence distracts from harms already occurring, such as errors, bias and misuse.

These are arguments and opinions, not experimental results. Skeptics have been wrong about specific milestones, and enthusiasts have been wrong about timing. Neither camp has a clean record.

What can today’s AI systems do and not do?

Current systems pass many professional-style exams, write working code, translate, summarize and hold fluent conversations. On creativity tests, GPT-4 reportedly outperformed 99 percent of humans in a 2023 evaluation. These are real capabilities.

They are also uneven. The same systems can fabricate facts, fail at tasks that seem simple to people, struggle with long-running goals and have limited ability to act in the physical world. Robotic efforts such as a humanoid robot learning to use a coffee machine in 2024 are demonstrations, not household-ready general machines. Whether the gaps close through scale, new architectures or both is the heart of the dispute.

What would count as strong evidence?

Anecdotes about a clever answer do not settle much. Stronger evidence would look like consistent performance on new tasks that the system was not trained on, over long stretches of work, with few of the basic mistakes that still appear today. It would also look like economic results: if a system could truly do most remote knowledge work, the effects would show up in labor and productivity data, not only in demonstrations.

That is the standard some commentators use to explain why the “AGI is here” claims have not changed much. The capabilities may be real, but general reliability across messy, open-ended jobs is a higher bar. The question of whether scaling current methods reaches that bar, or whether new ideas are needed, is the live technical dispute, and no one has shown the answer.

Why does it matter whether we call something AGI?

The label has consequences beyond semantics. Companies have used AGI as a mission statement; OpenAI’s charter describes ensuring AGI benefits humanity, and Meta’s chief executive named it a goal in January 2024. Contracts, funding, safety commitments and regulation sometimes hinge on definitions. A vague term lets each party claim success or failure as it suits them.

For readers, the practical advice is simple. When someone says AGI is coming in a certain year, ask which definition they mean, what evidence supports it and whether they have a stake in the answer.

Common questions

Has AGI already been achieved?

By some definitions, a few researchers say yes, and by stricter ones, most say no. There is no consensus, mainly because there is no agreed test.

Is ChatGPT an AGI?

It is a general-purpose language model that handles many tasks, but it is not widely accepted as AGI. DeepMind’s 2023 framework rated chatbots of that era as only “emerging” in generality, and current models still make basic errors.

What is the difference between AGI and superintelligence?

AGI usually means human-level breadth. Superintelligence means capability far beyond humans across nearly all domains. The second is a more speculative idea, and some argue it would follow quickly after the first while others doubt it.

When will AGI arrive?

Estimates range from a few years to many decades to never. Surveys and forecasting platforms have shifted earlier recently, but forecasters of every type have poor records on this question.

Is AGI dangerous?

Researchers disagree sharply. In 2023, leaders from DeepMind, OpenAI and Anthropic signed a statement that mitigating the risk of extinction from AI should be a global priority, while others consider such claims overstated. The risk assessment depends on assumptions that cannot yet be tested.

What to remember

AGI has no agreed definition, so any timeline is partly a statement about vocabulary. Recent forecasts are earlier than older ones, but the people making them have incentives and uneven track records. Current systems are impressive and uneven. The most reliable move is to ask what a claim means in concrete terms, and to treat confident dates with patience.

Further reading

A sober, well-regarded look at what AI can and cannot do. See it on Amazon (affiliate link)

As an Amazon Associate I earn from qualifying purchases. A commission may be earned if you buy through this link, at no extra cost to you.

Sources