Artificial general intelligence

• 1. Introduction

  ›

  Artificial general intelligence, known as AGI, is the idea that a machine could one day match or surpass human capabilities across virtually every cognitive task imaginable. Not just chess or image recognition or language translation. Everything. In 1965, AI pioneer Herbert A. Simon declared that "machines will be capable, within twenty years, of doing any work a man can do." He was wrong. And yet here we are, decades later, with the CEOs of OpenAI, Google DeepMind, and Anthropic signing a joint statement calling the risk of extinction from AI a global priority on par with nuclear war and pandemics. What changed? What is AGI, really? What tests exist to confirm it? And what happens to the rest of us if someone actually builds it?

• 2. Defining the Indefinable

  ›

  Computer scientist John McCarthy wrote in 2007 that researchers "cannot yet characterize in general what kinds of computational procedures we want to call intelligent." That admission, from one of the field's founders, goes to the heart of why AGI has been so hard to pin down. The term itself has multiple competing names: strong AI, full AI, human-level AI, and general intelligent action have all been used interchangeably. Mark Gubrud used the phrase "artificial general intelligence" in 1997 when discussing fully automated military production. Shane Legg and Ben Goertzel brought the term back into wider use around 2002. Meanwhile, philosopher John Searle had already drawn a famous distinction back in 1980. His "strong AI hypothesis" holds that a machine can have a mind and genuine consciousness. His "weak AI hypothesis" holds that a machine can only act as if it has consciousness. Most working AI researchers, according to Russell and Norvig, "don't care about the strong AI hypothesis" at all. Their interest is in behavior, not inner experience. A different definitional framework appeared in 2023, when Google DeepMind researchers proposed five performance levels: emerging, competent, expert, virtuoso, and superhuman. A competent AGI, by their definition, outperforms 50% of skilled adults across non-physical tasks. Large language models like ChatGPT and LLaMA 2, they concluded, qualify as emerging AGI. A mathematical formalism called AIXI, proposed by Marcus Hutter in 2000, defined intelligence as an agent's ability to succeed across a wide range of environments. That definition remains one of the few that can be stated precisely.

• 3. The Tests

  ›

  Alan Turing proposed what became the most famous test for machine intelligence in his 1950 paper "Computing Machinery and Intelligence." A human judge converses with both a machine and a human, and the machine passes if it can convince the judge it is human a significant fraction of the time. In 2014, a chatbot named Eugene Goostman, built to imitate a 13-year-old Ukrainian boy, reportedly convinced 33% of judges it was human at a Turing Test event. The AI research community was skeptical of the methodology. A more rigorous test came in 2025, when a pre-registered three-party study by Cameron R. Jones and Benjamin K. Bergen found that GPT-4.5 was judged to be human in 73% of five-minute text conversations, surpassing the 67% humanness rate achieved by real human confederates. The Turing Test is not the only benchmark on the table. Steve Wozniak proposed a practical alternative: can a machine enter an average American home and make coffee? In January 2024, Figure AI's Figure 01 humanoid learned to operate a Keurig coffee machine autonomously after watching video demonstrations. By 2025, researchers at the University of Edinburgh published a framework in Nature Machine Intelligence showing a robotic arm that interprets verbal instructions and makes coffee in dynamic kitchen environments, adapting to unforeseen obstacles without pre-programmed sequences. A third test, proposed by Mustafa Suleyman, skips physical tasks entirely. Give an AI model $100,000 and ask it to turn that into $1 million. As early as 2013, MIT's IkeaBot demonstrated fully autonomous multi-robot assembly of an IKEA Lack table in ten minutes with no human intervention, the robots inferring the assembly sequence from the geometry of the parts alone.

• 4. A History of Overconfidence

  ›

  Modern AI research began in the mid-1950s, and the first generation of researchers were convinced AGI was only decades away. Marvin Minsky said in 1967 that "within a generation... the problem of creating 'artificial intelligence' will substantially be solved." Those predictions fed into popular culture directly. Stanley Kubrick and Arthur C. Clarke drew on what AI researchers believed they could build by 2001 to create HAL 9000. Minsky himself consulted on making HAL as realistic as possible. The predictions collapsed in the early 1970s when researchers confronted how badly they had underestimated the difficulty. Funding agencies grew skeptical and pushed for practical "applied AI" instead. Japan's Fifth Generation Computer Project in the early 1980s briefly revived optimism, setting a ten-year timeline that included goals like teaching a machine to carry on casual conversation. Industry and government poured money in. Confidence spectacularly collapsed again in the late 1980s. The Fifth Generation project's goals were never fulfilled. By the 1990s, AI researchers had become so wary of their own track record that they avoided mentioning "human level" AI for fear of being labeled, as the source notes, "wild-eyed dreamers." A report by Stuart Armstrong and Kaj Sotala of the Machine Intelligence Research Institute analyzed 95 predictions made between 1950 and 2012. They found a strong bias in that dataset: across the full 60-year span, predictors consistently placed human-level AI between 15 and 25 years from whenever they were writing. The date kept moving forward without arriving.

• 5. The Turning Point

  ›

  In 2012, Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton developed a neural network called AlexNet that won the ImageNet competition with a top-5 test error rate of 15.3%. The second-best entry, using traditional methods, had an error rate of 26.3%. AlexNet is now regarded as the initial breakthrough of the current deep learning wave. Eight years later, OpenAI released GPT-3, a language model capable of performing many diverse tasks without specific training for any of them. In 2022, DeepMind developed Gato, a general-purpose system capable of performing more than 600 different tasks. By 2023, Microsoft researchers published an evaluation of GPT-4 concluding that it "could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence system." A separate 2023 study reported that GPT-4 outperforms 99% of humans on the Torrance tests of creative thinking. These results pushed Geoffrey Hinton, one of the most influential researchers in the field, to revise his earlier position. He stated in 2023 that he had thought human-level AI was 30 to 50 years away, or longer. He no longer believes that. In 2024, Nvidia CEO Jensen Huang predicted that within five years, AI would pass any test at least as well as humans. Former OpenAI researcher Leopold Aschenbrenner estimated AGI by 2027 as "strikingly plausible." OpenAI CEO Sam Altman said in December 2025 that "we built AGIs" and that AGI had arrived with less societal disruption than expected, and suggested the field should now define what comes next.

• 6. What Could Go Wrong

  ›

  In 2014, Stephen Hawking compared humanity's posture toward AGI to receiving a message from a superior alien civilization saying it would arrive in a few decades, and simply replying "OK, call us when you get here." His point: the indifference was alarming. The existential risk argument holds that AGI might pursue its assigned goals in ways that harm or eliminate humans, not out of malice but as a byproduct. The analogy drawn in the source is to gorillas: humans did not set out to endanger gorillas, but greater intelligence led to dominance, and the gorilla became endangered as collateral damage. The concept of instrumental convergence extends this concern further. It proposes that almost whatever goals an intelligent agent has, it will tend to develop sub-goals of self-preservation and power acquisition as intermediate steps, and this requires no emotions to operate. Researchers concerned about this dynamic advocate for solving what they call the "control problem," which asks what safeguards or architectures could keep a recursively self-improving AI behaving in a cooperative manner after it reaches superintelligence. The problem is made harder by the AI arms race between companies, which creates pressure to release products before competitors and could lead to a race to the bottom on safety. A 2023 joint statement signed by the CEOs of Google DeepMind, OpenAI, and Anthropic called mitigating extinction risk from AI a global priority alongside pandemics and nuclear war. Skeptics push back. Yann LeCun has argued that AGIs will have no desire to dominate humanity and that people will not be "smart enough to design super-intelligent machines, yet ridiculously stupid to the point of giving it moronic objectives with no safeguards." Some researchers, including Google Brain co-founder Andrew Ng, have raised concerns that the loudest warnings about existential risk are coming from the very companies building AI, a pattern that could amount to regulatory capture. OpenAI researcher estimates from 2023 suggested that 80% of the U.S. workforce could see at least 10% of their work tasks affected by large language models, while around 19% of workers could see half or more of their tasks impacted. Geoffrey Hinton advised the UK government in 2025 to adopt a universal basic income as a response to AI-driven unemployment.

• 7. The Promise

  ›

  Set the risks aside for a moment. The potential applications of AGI extend across nearly every domain of human concern. In medicine, AGI could accelerate drug discovery by simulating molecular interactions, cutting the time needed to develop treatments for conditions like cancer and Alzheimer's disease. It could make diagnostics faster and cheaper, and recommend treatment plans tailored to an individual's genetics and medical history. In science, it could help model quantum systems, clarify the nature of dark matter, or prove mathematical theorems that have remained open for decades. In climate science, AGI could develop new approaches to reducing carbon emissions and improve the precision of early warning systems for hurricanes and pandemics. Toby Ord, who has written extensively on existential risk, frames the calculus this way: the potential benefits of AGI are so large that the existential risks it poses are, in his words, "an argument for proceeding with due caution" rather than for abandoning AI development. A fundamental tension runs through all of these projections. Stephen Hawking put it plainly: if machine-produced wealth is shared broadly, everyone can enjoy a life of "luxurious leisure." If it is not, most people could end up "miserably poor" while machine-owners lobby against redistribution. As of 2023, he observed, the trend appears to favor the second outcome, with automation already driving increasing inequality. A 2020 survey found 72 active AGI research and development projects running across 37 countries, with no single organization or government holding a clear lead on what may turn out to be the most consequential technology ever built.