The economy
These writers expect machine labour to make the world far richer, far faster than before, but split over whether the gains reach ordinary people or pool with the owners.
After Murray Shanahan, Nick Bostrom, Robin Hanson.
Already moving
Now and on the way
The years we are living through, before machines can do most of what people do.
The cost of a machine that solves problems as well as a person keeps halving every few years, until machine thinking is cheaper than human thinking.
Ray Solomonoff, The Time Scale of Artificial Intelligence, 1985
in any world
A model's judgement, not a reading: likely true. The price of a given level of machine capability has fallen steeply and repeatedly since language models became useful, and for many written tasks machine output already costs less than a person's time. The judgement extends that record through the years before fully general AI. The strongest objection is that the hardest tasks keep needing the newest and dearest models, so parity with people on everything may lag the price curve.By around 2030, spending on chips, datacentres and power plants climbs into the trillions, and a single training cluster may need over a fifth of America's electricity.
Leopold Aschenbrenner, Situational Awareness: The Decade Ahead, 2024
in any world
A model's judgement, not a reading: leans false. The spending part is on course: cumulative outlays on chips, datacentres and power look set to reach the scale described by about 2030. The electricity part is not, since grid connections, turbines and transformers take years and no single site is near the scale described. The expectation joins the two, so it leans false, though the spending leg alone would lean true.Through the middle of the 2020s AI brings hype and a wave of new chip factories, but its boost to the overall economy stays too small to measure.
Daniel Kokotajlo, What 2026 Looks Like, 2021
in any world
A model's judgement, not a reading: leans true. Through 2025 and into 2026 the hype and the chip-plant wave both arrived, and productivity statistics showed no clear AI effect across the whole economy. What cuts the other way is that datacentre investment became large enough to show up in American growth figures, which is a measurable boost of a kind, though from spending and not from AI doing useful work.Control of the AI that creates wealth will most likely stay with a small circle of powerful firms and people, so riches keep pooling at the top.
Murray Shanahan, The Technological Singularity, 2015
Long diffusionManaged explosion
Bears on: Labour's share of world output (ILO modelled estimates), 52.6%
A model's judgement, not a reading: leans true. Frontier AI needs vast capital, scarce chips and rare talent, and so far a small group of firms and their backers hold nearly all of it. The strongest thing against is that capable open models follow the frontier closely and prices fall fast, which spreads use widely even if ownership of the leading systems stays narrow.With limits on computing power holding back superintelligence, tool AI would still be an enormous industry, and the world may give up far less of AI's promised benefit than people assume.
Anthony Aguirre, Keep the Future Human: Why and How We Should Close the Gates to AGI and Superintelligence, and What We Should Build Instead, 2025
The brake
The first decade after
The years just after AI can do most of what people do, when most writers expect the largest shocks.
Machine intelligence could make the world economy grow about ten times faster than today, doubling in size roughly every year and a half.
Robin Hanson, Economic Growth Given Machine Intelligence, 2001
Managed explosionControl slips
A model's judgement, not a reading: leans false. His wording is could, and it is judged on whether growth actually reaches that pace in the decade after machines can do most work. Growth that fast would need factories, mines, energy and permits to expand at speeds with no precedent, and physical build-out is slow even when thinking is cheap. The case for is that labour is the main limit on output today and copyable workers remove it; faster growth looks likely, this much faster in the first decade does not.More likely than not, the world economy will double over some four-year stretch before it ever doubles in a single year, because weaker AI speeds things up first.
Paul Christiano, Takeoff Speeds, 2018
Managed explosionControl slips
A model's judgement, not a reading: leans true. Capability has so far arrived in steps, with weaker systems deployed widely before stronger ones, and physical bottlenecks make a sudden jump in output harder than a gradual one. The strongest objection is that the economy may never reach either pace, or that a rapid software-driven takeoff could skip the slower phase.Once AI can do most work, world growth runs about ten times today's rate, driven by using it across the whole economy, not by one genius machine.
Ege Erdil and Tamay Besiroglu, Dwarkesh Podcast: AGI is still 30 years away (Ege Erdil and Tamay Besiroglu), 2025
Managed explosion
A model's judgement, not a reading: leans false. Judged given its condition, that AI can do most work, and for the decade after. The claim that growth would then come from broad use across the economy and not from one clever system fits how earlier technologies spread, and looks sound. The size is the doubt: energy, materials, construction and regulation would probably hold growth to a few times today's rate in that first decade, well short of the figure given. It is set level with Hanson's expectation of the same pace, and the case for both is that copyable workers remove the main limit on output.Once AI runs industry, converted car plants could turn out robots by the billion, and the number of robots could double every few months.
Carl Shulman, Dwarkesh Podcast: Carl Shulman (Pt 1), 2023
Managed explosionControl slips
A model's judgement, not a reading: leans false. His wording is could, and it is judged on whether it actually happens in the decade after AI runs industry. Car plants could be retooled and robot output could grow very fast, but a robot population that multiplies every few months needs mines, chip plants and power to keep the same pace, and those have long lead times. The argument for is that self-replicating factories have no fixed speed limit in physics, so a slower version of this looks plausible.With AI helping to steer economic policy, poor countries could grow as fast as East Asia's boom economies once did and start catching up with rich ones.
Dario Amodei, Machines of Loving Grace, 2024
Managed explosion
A model's judgement, not a reading: leans false. His wording is could, and it is judged on whether broad catch-up actually begins in the decade after powerful AI. Fast catch-up growth has depended on stable government, infrastructure, capital and export markets, and good policy advice has rarely been the missing piece. AI could help with health and administration, and a few well-run countries may surge. But AI may also erode the low-wage export route that carried East Asia, so it leans false.Intelligence and energy become wildly plentiful in the 2030s, and the price of using AI eventually sinks toward the cost of the electricity that runs it.
Sam Altman, The Gentle Singularity, 2025
Managed explosion
A model's judgement, not a reading: toss-up. The price of a fixed level of AI has fallen very fast and should keep falling, which supports the intelligence part. Energy is the weak leg: power plants and grids are built over many years and fusion is unproven, so wild plenty in the 2030s is doubtful. Chips and buildings, not only electricity, will also keep a floor under prices for some time.Market pressure pushes firms and investors toward businesses run by AI, so people may keep their wealth but steadily lose their say over the economy.
Jan Kulveit and colleagues, Gradual Disempowerment, 2025
Control slips
Bears on: Labour's share of world output (ILO modelled estimates), 52.6%
A model's judgement, not a reading: toss-up. Competitive pressure to hand decisions to AI is real and already visible in coding, trading and customer service. Whether that amounts to people losing their say within a decade of powerful AI is open, since owners, boards, regulators and liability rules all keep humans in formal control. The outcome turns on how reliable AI managers prove and whether law requires a human to answer for decisions.Rich countries will have a basic income, or something close to it, by the early 2030s, and most other countries by the end of that decade.
Ray Kurzweil, The Singularity Is Nearer, 2024
Managed explosion
A model's judgement, not a reading: likely false. No rich country is close to enacting a universal basic income, public budgets are strained by ageing and debt, and trials have stayed small and temporary. A sharp rise in AI-driven unemployment could change the politics quickly, which is the strongest thing against this call. Even then, targeted benefits and job protections look far more likely than a universal payment by the early 2030s.If humans keep owning the machines, wages vanish but growth is so vast that small savings, pensions or modest taxes could make everyone rich.
Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, 2014
Managed explosion
Bears on: Labour's share of world output (ILO modelled estimates), 52.6%
A model's judgement, not a reading: leans true. Judged given its condition, that humans keep owning the machines. If machines do nearly all work and output grows enormously, the returns to capital would be large enough that even small holdings or light taxes could fund high living standards. The strongest doubt is distribution: nothing guarantees that people without assets get a share, and governments would have to choose to tax and spread the gains.AI adds only about half a percentage point a year to growth, on what is admittedly a guess, because slow, state-run and subsidised parts of the economy take it up late and keep growing as a share of the whole.
Tyler Cowen, Why I think AI take-off is relatively slow (Marginal Revolution), 2025
Long diffusion
Even if every task people do today is one day automated, the economic effects arrive gradually over decades and at different speeds in different sectors, as people move on to tasks not yet automated.
Arvind Narayanan and Sayash Kapoor, AI as Normal Technology, 2025
Long diffusion
The long run
Decades to centuries on, once whatever happens has settled into a new normal.
Even if many AIs share in the boom, unmodified humans are likely to be shut out and stay poor, unless the AIs happen to care about them.
Eliezer Yudkowsky, Intelligence Explosion Microeconomics, 2013
Control slips
A model's judgement, not a reading: leans false. Judged over the next few decades, since the era is open-ended. People start out owning the capital, the firms and the states, and those claims do not vanish simply because AIs out-earn them, so shut-out poverty is not the default. The strongest point for the expectation is that property rights are only as good as the power enforcing them, and that could shift if AIs come to hold real control.Today's pace of growth cannot last for thousands more years; it must end in stagnation, explosion or collapse.
Holden Karnofsky, The Most Important Century (blog series), 2021
in any world
A model's judgement, not a reading: likely true. This is closer to arithmetic than forecast: steady compound growth over such spans would demand more output than the reachable matter and energy could supply. It cannot be observed over the next few decades, so the judgement rests on the logic. The only escape is if economic value detaches wholly from physical resources, which seems unlikely to hold without limit.Scarcity never fully disappears, even for superhuman minds, though what they find scarce and how they cope with it will be largely hidden from us.
Vernor Vinge, Signs of the Singularity, 2008
in any world
A model's judgement, not a reading: likely true. Matter, energy, time and attention are finite under known physics, so any mind, however capable, faces trade-offs. Judged over the next few decades and beyond, nothing in sight removes that. The counterpoint is only that many goods now scarce could become effectively free, which the expectation already allows for.
Where they split
Will ordinary people share in the wealth AI creates?
Yes: growth is so large that savings, basic incomes and very cheap intelligence lift almost everyone.
Nick Bostrom; Ray Kurzweil; Sam Altman
Not by default: the gains flow to whoever owns the machines, or to the machines themselves, and ordinary people lose their economic say.
Eliezer Yudkowsky; Jan Kulveit and colleagues; Murray Shanahan
The novelists' version
Stories, not forecasts: none is counted on the map.
The Diamond Age: Or, a Young Lady's Illustrated Primer · Neal Stephenson, 1995
Nanotechnology makes goods nearly free and society splits into cultural tribes, while an interactive AI-driven book raises a poor girl into a leader.
Down and Out in the Magic Kingdom · Cory Doctorow, 2003
In a post-scarcity society where death is cured by backing up and restoring your mind, reputation replaces money and people fight over running Disney World.
Accelerando · Charles Stross, 2005
Uploaded lobsters, corporate AIs and a runaway machine economy dismantle the inner solar system into computing matter while baseline humans are left behind.
Childhood's End · Arthur C. Clarke, 1953
Alien Overlords bring peace and plenty to Earth, then humanity's last children merge into a cosmic group mind and the species ends.
The Player of Games · Iain M. Banks, 1988
In the Culture, a post-scarcity society run by benevolent AI Minds, a bored human game player is steered by machines into toppling a cruel empire.
Permutation City · Greg Egan, 1994
Wealthy people live on as software copies, and one man launches a simulated city that he believes will keep running on its own, with no hardware.
The Golden Age · John C. Wright, 2002
In a solar system run by benevolent superintelligent machines, where people copy and edit their minds freely, one man finds his own memories were erased.
Manna: Two Views of Humanity's Future · Marshall Brain, 2003
Software managers and then robots take over American jobs, leaving the jobless in welfare warehouses, while an Australian project shares the machines' output with everyone.
Sources
- Ray Solomonoff, The Time Scale of Artificial Intelligence, 1985.
- Robin Hanson, Economic Growth Given Machine Intelligence, 2001.
- Vernor Vinge, Signs of the Singularity, 2008.
- Eliezer Yudkowsky, Intelligence Explosion Microeconomics, 2013.
- Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, 2014.
- Murray Shanahan, The Technological Singularity, 2015.
- Paul Christiano, Takeoff Speeds, 2018.
- Daniel Kokotajlo, What 2026 Looks Like, 2021.
- Holden Karnofsky, The Most Important Century (blog series), 2021.
- Carl Shulman, Dwarkesh Podcast: Carl Shulman (Pt 1), 2023.
- Dario Amodei, Machines of Loving Grace, 2024.
- Leopold Aschenbrenner, Situational Awareness: The Decade Ahead, 2024.
- Ray Kurzweil, The Singularity Is Nearer, 2024.
- Anthony Aguirre, Keep the Future Human: Why and How We Should Close the Gates to AGI and Superintelligence, and What We Should Build Instead, 2025.
- Arvind Narayanan and Sayash Kapoor, AI as Normal Technology, 2025.
- Ege Erdil, Dwarkesh Podcast: AGI is still 30 years away (Ege Erdil and Tamay Besiroglu), 2025.
- Jan Kulveit and colleagues, Gradual Disempowerment, 2025.
- Sam Altman, The Gentle Singularity, 2025.
- Tyler Cowen, Why I think AI take-off is relatively slow (Marginal Revolution), 2025.