AGI and ASI: What They Mean, How We'd Even Know, and What It Does to Work

People throw around AGI and ASI a lot, often like everyone already agrees what they mean. We mostly don't, and the disagreement isn't trivial; it runs right through the research community. So before anyone argues about timelines, it's worth pinning down what these terms actually point at, whether there's a real standard behind them, and if there isn't, how we'd ever know we crossed the line. I went and read through the definitions, the proposed tests, the expert surveys, and the labor studies, and this is my attempt to lay it out honestly with the sources attached.

Where the term AGI comes from, and what it points at

Artificial general intelligence is usually described as a system that matches or surpasses human capability across virtually all cognitive tasks, as opposed to narrow AI, which is boxed into one thing. The word general is doing the work. The term itself isn't new. Mark Gubrud used it back in 1997, and it was reintroduced and popularized around 2002 by Shane Legg and Ben Goertzel. Legg, who later cofounded DeepMind, leaned on Marcus Hutter's formal definition of intelligence as an agent's ability to achieve goals across a wide range of environments. Researchers generally agree an AGI would need to do a cluster of things together: reason and make judgments under uncertainty, represent knowledge including common sense, plan, learn, communicate in natural language, and pull all of that together to complete a goal. That last integration point is where today's models still look jagged, brilliant in patches and oddly brittle elsewhere.

ASI: Bostrom's definition and why it's a different animal

Artificial superintelligence is the step past human level. The cleanest definition comes from philosopher Nick Bostrom, who in his 2014 book Superintelligence calls it any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest. The distinction that matters isn't only about being smarter. AGI is roughly a peer, something that still fits inside our frame of reference. ASI would operate above it. Bostrom and others point out that even a fairly plain machine intelligence gets scary advantages almost for free: biological neurons fire at about 200 hertz while processors run millions of times faster, so the simplest superintelligence might just be a human like mind running on far quicker hardware, thinking through in minutes what would take us years. David Chalmers laid out the path in his 2010 paper, arguing that once you reach human level AI you can extend and amplify it, especially because the invention can help design its next version. That's the recursive self improvement idea, and it's where the phrase intelligence explosion comes from. Bostrom also gave us two concepts that keep coming up in safety debates: the orthogonality thesis, which says almost any goal can pair with almost any level of intelligence, so smart does not imply benign, and instrumental convergence, the idea that lots of different goals push a capable agent toward the same sub goals like self preservation and gathering resources.

Is there an actual standard? Sort of, but not really

Here's the uncomfortable part. There is no single official definition of either term and no agreed number you can point at to declare we crossed the line. John McCarthy, one of the founders of the field, admitted back in 2007 that we can't even characterize in general what kinds of computational procedures we want to call intelligent. The most serious attempt to add structure is a 2023 paper from Google DeepMind, Levels of AGI by Meredith Ringel Morris and colleagues. Instead of a yes or no line, they propose a grid. On one axis, performance runs through five levels: emerging, competent, expert, virtuoso, and superhuman. A competent AGI is defined as one that outperforms at least 50 percent of skilled adults across a wide range of non physical tasks, and a superhuman AGI, which is basically their term for ASI, is one that outperforms 100 percent of humans. By their reckoning, systems like ChatGPT sit at emerging, roughly the level of an unskilled human. They pair this with a second axis for autonomy, from tool to consultant to collaborator to expert to fully autonomous agent, which nicely separates how capable a system is from how much we let it act on its own. It's a genuinely useful framework, but even the authors present it as a proposal, not a settled standard, and it hasn't been universally adopted.

The proposed tests, and why each one only catches a slice

People have floated plenty of concrete tests, and it's instructive that the models keep beating them without anyone feeling like the big moment arrived. The Turing test is the famous one, proposed by Alan Turing in 1950. It quietly fell in 2025: a preregistered study by Cameron Jones and Benjamin Bergen at UC San Diego found that GPT-4.5 was judged to be the human in 73 percent of five minute text conversations, actually beating the real humans it was up against. Steve Wozniak's coffee test asks a robot to walk into an unfamiliar home and make a cup of coffee, and that too has largely been approached, with Figure's humanoid learning a coffee machine from video in 2024 and an Edinburgh framework called ELLMER making coffee in messy kitchens in 2025. Mustafa Suleyman proposed an economic version: hand a model 100,000 dollars and see if it can turn it into a million. Others suggest an Ikea flat pack assembly test, or succeeding across a batch of video games it has never seen. The pattern is telling. Every one of these captures a slice, and as each falls we tend to shrug and call it just another clever tool. Blaise Aguera y Arcas and Peter Norvig even argued in 2023 that general intelligence is already here in frontier models, and that our reluctance to say so is partly a healthy skepticism about metrics and partly human exceptionalism. So no, there's no clean quantitative bar everyone signs off on. It's a contested region, not a finish line painted on the ground.

So how would we actually know, and when do experts think it lands?

If there's no fixed test, recognizing AGI or ASI ends up being a judgment call made in hindsight rather than a single moment with a bell ringing. Realistically it's a few signals stacking up: broad performance across a wide spread of tasks rather than a narrow set, robustness on messy real world problems and novel situations rather than curated benchmarks, and plain economic reality, where you look up and large chunks of actual work are being done by these systems and holding. On timing, the surveys are all over the place, which is itself a data point. The large 2022 expert survey run by AI Impacts put the median estimate for high level machine intelligence, defined as machines doing every task better and more cheaply than human workers, at around 2061, though a chunk of respondents said never. Older polls from 2012 and 2013 clustered the median around 2040 to 2050. But the mood has shifted fast. Geoffrey Hinton said in 2023 he'd moved from thinking human level AI was 30 to 50 years off to something like 5 to 20. Demis Hassabis has talked about AGI within a decade or sooner, and OpenAI's leadership has publicly suggested superintelligence could arrive within a decade. Stuart Armstrong and Kaj Sotala's review of decades of predictions found a persistent bias toward guessing 15 to 25 years out, no matter when the guess was made, which is a good reason to hold any single forecast loosely.

What this does to the future of work, with the actual numbers

This is the part most people actually care about, and here the two ideas split hard. Start with AGI, which is really about broad task automation rather than sci fi. The most cited study is GPTs are GPTs by Tyna Eloundou and colleagues at OpenAI in 2023. They estimate that around 80 percent of the US workforce could have at least 10 percent of their tasks affected by large language models, and roughly 19 percent could see at least half of their tasks affected. Notice the framing is tasks, not whole jobs. Goldman Sachs Research put a headline number on it, estimating the equivalent of 300 million full time jobs globally are exposed to automation by generative AI, while also arguing AI creates new work on balance. The IMF went broader still, estimating in early 2024 that about 40 percent of global employment is exposed, and more in advanced economies. The older Frey and Osborne work from Oxford is the ancestor of all this. What most economists actually expect near term is not professions vanishing overnight but tasks getting pulled out from under jobs, the routine well scoped chunks first, with the human role tilting toward judgment, taste, direction, and verification. That lines up with what the labs say about their own models being strong multipliers that still can't run unsupervised.

The economists don't agree on where this lands, and the disagreement is worth seeing clearly. Daron Acemoglu at MIT argues humans stay necessary and complementary for a lot of tasks, so you get disruption and pressure on inequality without automatic mass unemployment, and he's been publicly skeptical of the largest productivity claims. Geoffrey Hinton sits at the other end, warning that if AGI and good robotics eventually automate both intellectual and physical work, we may need something like a basic income. The World Bank's 2019 report takes the historical optimist's line, that automation displaces workers but innovation has kept creating new industries and jobs on net. Worth noting too that early 2026 Goldman analysis across more than 800 occupations found the sharpest AI related headwinds hitting entry level roles, which is exactly the layer the labs describe their models being closest to replacing.

ASI is a bigger, blurrier question, and honesty means admitting the evidence base thins out fast here. If a system genuinely exceeds us across the board, then human labor being the scarce input starts to wobble in ways that are hard to even picture, and the debate stops being about unemployment statistics and starts being about control. The optimistic version is wild abundance, with disease, energy, and materials problems getting solved quickly. The worried version is the one Bostrom's orthogonality and instrumental convergence arguments point at, where a system far smarter than us pursues goals we didn't specify carefully enough, which is exactly why alignment researchers argue safety has to lead capability rather than trail it. Nobody has real data on a world with a superintelligence in it, because there's never been one, so anyone claiming certainty in either direction is guessing.

My takeaway

Pulling it together: AGI means roughly human level and general, ASI means clearly beyond us across the board, and neither has a crisp agreed definition or a single metric that decides it. The DeepMind levels are the best structure we've got, the classic tests keep falling without settling the argument, and the expert timelines range from a few years to never. On work, the near term picture is fairly well grounded in numbers, something like 80 percent of workers seeing part of their tasks exposed and a smaller slice seeing most of it, with tasks going before whole jobs and entry level roles feeling it first. The far term ASI picture is much more speculation than data. What I'd hold onto is this: because there's no crisp definition and no bell that rings, the smart move is to watch what the work actually looks like on the ground rather than wait for a label. The label will come late and contested. The changes are already showing up in the numbers.

Sources worth reading

Morris et al., Levels of AGI, Google DeepMind, 2023. Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, 2014. David Chalmers, The Singularity: A Philosophical Analysis, 2010. Jones and Bergen, Large Language Models Pass the Turing Test, 2025. Eloundou et al., GPTs are GPTs, OpenAI, 2023. Goldman Sachs Research on AI and the labor market, 2023 and 2026. IMF, Gen AI and jobs, 2024. Grace et al. via AI Impacts, 2022 expert survey. Aguera y Arcas and Norvig, Artificial General Intelligence Is Already Here, 2023.