In 2020, the median forecaster on Metaculus, a platform that has tracked expert and public predictions on hard questions for over a decade, put artificial general intelligence roughly fifty years away, somewhere around 2070. By February 2026, that same community's median had compressed to 2033, with a 25% chance assigned to 2029. That is not a small revision. That is the fastest a serious technological forecast has moved in modern memory, and it happened without any single dramatic breakthrough, just a steady accumulation of capability gains that kept beating expectations. Before asking what AGI would do to the economy or to society, it is worth sitting with how genuinely unsettled the timeline question still is, because almost everything that follows in this piece depends on an answer nobody actually has.
Chapter 1What is AGI, and why do experts disagree by decades?
Artificial general intelligence, in its most common working definition, means a system that can do most cognitive work at human level or better across nearly any domain, not just the narrow task it was trained on. The disagreement about when it arrives is not really a disagreement about the technology's trajectory. It is a disagreement about the definition itself, and that matters more than it sounds.
On one end, Dario Amodei of Anthropic and Sam Altman of OpenAI have both pointed to windows as near as 2026 or 2027 for AI systems broadly better than humans at most tasks. Elon Musk has offered similarly short timelines. On the other end, Demis Hassabis of Google DeepMind puts roughly 50% odds on 2028 to 2030, while former OpenAI researcher Andrej Karpathy places genuine AGI a full decade out, and Yann LeCun and Gary Marcus argue current model architectures cannot reach it at all, regardless of how much longer you wait. Sam Altman himself has called AGI "not a super useful term" precisely because the goalposts move every time a model clears the last one.
Prediction markets add a useful, if messy, second data point. Kalshi traders assigned roughly 40 to 55% probability to OpenAI achieving AGI by 2030 in early 2026, while Polymarket priced the chance of AGI by 2027 at only 9 to 14%. Samotsvety Forecasting, a superforecaster group with a genuinely strong track record on other hard questions, put the probability at roughly 28% by 2030 in its January 2026 update, itself a sharp compression from a 32% chance within twenty years back in 2022. A 2023 survey of 2,778 published AI researchers landed on a far more conservative 50% probability of high level machine intelligence by 2040. The honest summary is that the people building these systems tend to forecast sooner, and the people studying them from further outside tend to forecast later, and the gap between those two groups has not closed even as both groups have moved their own estimates earlier.
Chapter 2Is the singularity the same claim as AGI, or something bigger?
These two words get used almost interchangeably in casual conversation, and they should not be. AGI describes a system reaching human level general capability. The singularity describes something that happens after that, and it is a much larger and much more uncertain claim.
The idea traces to a specific 1965 essay by mathematician I. J. Good, who argued that an "ultraintelligent machine" capable of designing better machines than any human would trigger a feedback loop, each generation of machine improving the next faster than the last, producing what he called an intelligence explosion that would leave human intelligence "far behind." Vernor Vinge popularised the term singularity for this scenario in the 1980s, borrowing the physics metaphor of a point beyond which the normal rules stop applying and outcomes become fundamentally unpredictable to anyone on the near side of it. Ray Kurzweil's 2005 book gave the idea its widest public audience, predicting the singularity by 2045 based on extrapolating computing power growth, and arguing that the distinction between human and machine intelligence would eventually disappear entirely.
This distinction matters for one practical reason. AGI is, at least in principle, a measurable engineering milestone, however contested the definition. The singularity is a claim about what happens after that milestone, involving a self improving system whose subsequent behaviour is by definition difficult for humans to predict or fully understand, which is a fundamentally different, and much harder to verify, kind of claim. Geoffrey Hinton, one of the field's most senior researchers, revised his own timeline for transformative AI from fifty years to five to twenty, and separately assigns something like a 10 to 20% probability to AI causing human extinction, a figure that reflects the singularity style concern about loss of control, not merely the AGI style concern about job displacement. Treating these as the same question, or assuming that solving the economic disruption question also handles the control question, is one of the more common errors in how this topic gets discussed publicly.
Chapter 3What do the actual economic models say, and why do they disagree by a factor of twenty?
This is where the disagreement stops being philosophical and becomes something you can actually put a number on, and the range of numbers on the table is remarkable.
Daron Acemoglu, the MIT economist who shared the 2024 Nobel Prize in Economic Sciences, published an estimate in 2024 that AI would lift US GDP by roughly 1.1 to 1.6% over a decade, calling this a "nontrivial, but modest" effect. His reasoning is specific rather than reflexively sceptical. He estimates that AI technology could touch around 20% of US labour market tasks, but only about a quarter of those, or 5% of the total economy, can be automated profitably within a decade once implementation costs are weighed against the benefit.
Goldman Sachs Research sits at the opposite end, forecasting that generative AI could lift global GDP by 7%, worth roughly 7 trillion dollars, over ten years, and could raise labour productivity in developed markets by around 15% once fully adopted. McKinsey Global Institute's estimate is higher still, projecting an additional 13 trillion dollars in global economic activity by 2030, equivalent to roughly 16% higher cumulative GDP compared to today.
The disagreement is not really about whether AI works. It is about two specific assumptions. First, how many real world tasks actually get automated once you account for implementation cost, error tolerance, and the messy, multi step nature of most real jobs, which is Acemoglu's central sticking point. Second, whether automation mainly replaces existing tasks at lower cost, which is roughly all Acemoglu's model counts, or whether it also creates genuinely new tasks and reallocates displaced workers into them, an effect Goldman explicitly builds into its estimate and which Acemoglu's own past research, on what he calls the reinstatement effect, argues has always determined whether past general purpose technologies like electricity or the internet actually raised living standards broadly or mostly just displaced people.
Chapter 4What does this actually look like in jobs data that already exists, not just forecasts?
Some of this has stopped being purely hypothetical, which is worth separating from the parts that remain genuinely speculative.
The World Economic Forum's Future of Jobs Report 2025, surveying over 1,000 employers representing 14 million workers across 55 economies, projects that 92 million jobs will be displaced globally by 2030 while 170 million new ones are created, a net gain of 78 million, though the report is explicit that the people losing jobs and the people gaining new ones are often not the same individuals, which is where the real disruption sits, hidden inside a headline number that looks reassuring.
Goldman Sachs Research's baseline case estimates 6 to 7% of the US workforce will be displaced over a roughly ten year adoption window, with a plausible range of 3 to 14% depending on how fast adoption actually moves, translating to a manageable 0.6 percentage point rise in unemployment if spread evenly across a decade, but a much sharper shock if the transition front loads. Some of this is already visible in the data rather than projected. Unemployment among 20 to 30 year olds in tech exposed occupations has risen almost 3 percentage points since early 2025, faster than for their peers in other trades, and McKinsey's own 2025 survey found 51% of organisations already reducing entry level hiring specifically because of generative AI, not five years from now, already. IBM announced in 2025 it would eliminate roughly 7,800 back office roles within five years, citing AI directly. None of this is the mass unemployment scenario some warn about, but it is also not nothing, and it is happening well before anything resembling AGI, let alone the singularity, has arrived by any measure.
Chapter 5If it does arrive at scale, what happens to the social fabric, not just employment numbers?
This is the part of the question that has the least hard data behind it, and it deserves to be treated with real caution rather than confident prediction, because unlike GDP forecasts, there is no dataset of prior AGI arrivals to check any of this against.
The closest historical parallel economists reach for is not a previous piece of software, it is the agricultural and industrial revolutions, both of which economist Robin Hanson has described as prior "singularities" in the sense that each one multiplied the underlying rate of economic growth by somewhere between 60 and 250 times compared to what preceded it, while also completely restructuring where people lived, what a normal working day looked like, and what family and community structures were built around. Those transitions delivered enormous long run gains in living standards, and also generations of genuine social upheaval, urbanisation shock, and inequality before those gains were broadly shared, a pattern Acemoglu's own research on past general purpose technologies argues was never automatic and always required specific policy and labour power to force a fairer distribution of the gains.
Universal basic income is the policy response that gets discussed most often as a stabiliser, and it has actually been tested, at a scale far smaller than a full AGI transition would demand. Finland's 2017 to 2018 trial, giving a fixed monthly payment to 2,000 unemployed citizens with no work requirement, found meaningful improvements in reported wellbeing and mental health, but no significant change in employment outcomes relative to the control group, a genuinely mixed result that both UBI advocates and sceptics have since cited selectively. Longer running studies from GiveDirectly's basic income programme in Kenya have found broadly similar patterns, real wellbeing gains, modest and debated effects on work itself. What none of these trials tested, because none of them could, is what a UBI style policy would need to look like if the disruption were not a few thousand participants but a meaningful share of an entire national workforce at once, which is the actual scenario an AGI level transition would raise.
Chapter 6Does this make things simpler or more complicated for humanity, and which is it actually?
Two entirely coherent stories exist side by side here, and the honest answer is that the technology itself does not determine which one plays out.
The optimistic version, sometimes called an age of abundance, argues that if AGI genuinely automates most cognitive labour, the cost of goods and services should fall dramatically, freeing human time for whatever people actually want to do rather than what they need to do to earn a living, echoing the earlier hopeful predictions made about automation during the industrial revolution that eventually did materialise, for most people, over a much longer timeframe than anyone initially expected.
The more complicated version starts from Acemoglu's own observation, buried inside his otherwise modest projections, that GDP could rise while overall welfare declines, if the gains concentrate in whoever owns the compute, the models, and the capital rather than flowing to displaced workers through new jobs or redistribution. This is not a technology problem. It is the same distributional question every previous general purpose technology has raised, except compressed into a potentially much shorter timeframe and applied to a much wider slice of the workforce than steam power or electricity ever touched at once, since those transformed specific industries over decades while a genuine AGI transition could touch cognitive work broadly and simultaneously.
Chapter 7So, where does this actually leave the question?
The timeline is genuinely unknown, compressing fast enough that a 2033 median forecast in 2026 could look as outdated as a 2070 forecast looked six years ago, or could turn out to be too aggressive, and nobody currently forecasting has a track record long enough to say which. The economic models disagree by a factor of twenty not because economists cannot do arithmetic, but because the actual crux, how many tasks genuinely get automated and whether new ones get created fast enough to absorb displaced workers, has never been settled for any past technology until decades after the fact. What is not actually in question is that a meaningful slice of this disruption is already showing up in hiring data for young, tech exposed workers, years before anything resembling full AGI. The single thing every serious economist across this entire spectrum, from Acemoglu to Goldman Sachs, agrees on without exception is that the outcome for ordinary people depends almost entirely on policy and distribution choices made during the transition, not on the underlying capability of the technology itself. That is the one part of this story that is not actually hypothetical.