Work

Most of these writers expect machines to climb from routine tasks to skilled ones, with pay holding up for a while before human labour loses most of its value.

After Dario Amodei, Hans Moravec, Robin Hanson.

The forge and the scriptorium. An illustration made in v0, not evidence.

Already moving

Labour's share of US business output, change on a year before

-3.4% · 2026

Employment of young workers against older ones in the most AI-exposed jobs, gap in index points

-17.4 index points · 2026

Share of US work hours done with generative AI

5.3% · 2026

Now and on the way

The years we are living through, before machines can do most of what people do.

Machines first take the skills evolution never honed in us, like calculating and filing, then reach walking, handwork and social life within half a century at today's pace.
Hans Moravec, When will computer hardware match the human brain?, 1998
Long diffusion
A model's judgement, not a reading: leans true. The order has held: calculation and filing went first, and language, images and code followed. Walking robots now work in demonstrations and pilots, and social conversation with machines is common, with roughly two decades of his window left. Skilled handwork in cluttered real settings is the laggard and the strongest reason for doubt.
Automation climbs into ever more skilled jobs, until truly productive work belongs to a shrinking elite and technological unemployment finally arrives.
Vernor Vinge, The Coming Technological Singularity, 1993
in any world
A model's judgement, not a reading: leans false. Automation is climbing into skilled work, as the first part says. But through 2026 unemployment in rich countries stayed near ordinary levels, and in the period before fully general AI new tasks and slow adoption should keep most people employed. Early weakness in entry-level jobs exposed to AI is the strongest sign the other way.
As it has for two centuries, automation keeps removing jobs at the bottom of the skills ladder and adding better-paid ones at the top, so education spending climbs.
Ray Kurzweil, The Law of Accelerating Returns, 2001
in any world
A model's judgement, not a reading: toss-up. This held for two centuries, and continuity is usually the safer bet. But current AI bites first on routine cognitive and entry-level professional work, while much low-paid manual work is untouched, which breaks the bottom-up pattern. Whether more schooling still pays as it did is now an open question.
In their central story, by late 2026 AI starts taking jobs, above all junior programming work, while people who can manage teams of AI earn a fortune.
The AI 2027 authors, AI 2027, 2025
in any world
A model's judgement, not a reading: leans true. By 2026 several studies and hiring figures pointed to weaker demand for junior programmers and other young workers in AI-exposed roles, while pay for scarce AI talent rose to extraordinary levels. The strongest objection is that higher interest rates and the unwinding of earlier over-hiring explain much of the junior slump, and that economy-wide job losses from AI remain hard to see.
By about 2027, AI agents work like remote colleagues who can fully replace a software engineer or any other computer-based worker.
Leopold Aschenbrenner, Dwarkesh Podcast: Leopold Aschenbrenner, 2027 AGI and the return of history, 2024
Managed explosionControl slips
A model's judgement, not a reading: leans false. Agents in 2026 can complete long, well-specified software tasks, and progress has been fast. Fully replacing a software engineer, or any computer-based worker, also requires reliability over weeks, judgement in ambiguous situations and accountability, and those gaps look unlikely to close everywhere by 2027. Continued rapid gains in the length of tasks agents can finish is what cuts the other way.
An AI able to do every job that can be done remotely is still about two decades away, arriving around 2045.
Ege Erdil, Dwarkesh Podcast: AGI is still 30 years away (Ege Erdil and Tamay Besiroglu), 2025
Long diffusion
A model's judgement, not a reading: leans false. The rate at which agents are taking on longer and harder computer tasks suggests full coverage of remote work sooner than two decades out, perhaps well within the 2030s. The case for a later date is strong, though: the last tasks are the hardest, reliability improves slowly, and forecasts of this kind have slipped before.

The first decade after

The years just after AI can do most of what people do, when most writers expect the largest shocks.

In his model, once machines can do most jobs, human wages eventually fall almost as fast as computer prices, possibly below what a person needs to live.
Robin Hanson, Economic Growth Given Machine Intelligence, 2001
Managed explosionControl slips
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 machines can do most jobs. If a machine substitute is available for most work and its cost keeps falling, competition drags wages for that work down with it. What cuts against is that tasks machines cannot do, or are not allowed to do, would become more valuable, and higher output could raise some human wages for a long transition.
When copyable machines can do nearly every job more cheaply, wages sink below subsistence, as happened to horses, except where buyers prefer human-made work.
Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, 2014
in any world
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 copyable machines do nearly every job more cheaply. On that premise the market wage for such work falls toward the machine's running cost, which could sit below a person's needs. The strongest objection is that people, unlike horses, vote and own property, so minimum wages, transfers and protected roles would stop pay from being the thing people live on.
For a while people stay employed and grow more productive, because they keep an edge on some tasks, especially in the physical world.
Dario Amodei, Machines of Loving Grace, 2024
Managed explosionLong diffusion
A model's judgement, not a reading: likely true. This matches what is visible: AI is being used mainly alongside workers, and robots lag software by a wide margin in the physical world. The claim is modest, saying only that this holds for a while. The risk to it is a faster jump in robotics or in agent reliability than recent progress suggests.
Whole kinds of jobs disappear, but the world grows rich fast enough to try bold new policies, and people invent new work that feels important.
Sam Altman, The Gentle Singularity, 2025
Managed explosion
A model's judgement, not a reading: toss-up. That whole kinds of jobs disappear is very likely, and people have always invented new work they find important. The doubtful part is the middle: wealth may arrive unevenly, and bold policy tends to trail events, so the world may not grow rich fast enough to cushion the losses. The expectation joins all three parts, and the weakest of them is open.
Most likely, new kinds of jobs will not appear fast enough or in large enough numbers to replace those AI takes away over the next couple of decades.
Mustafa Suleyman, The Coming Wave, 2023
Long diffusion
A model's judgement, not a reading: toss-up. History favours new work appearing, but past machines did not match people across nearly all thinking tasks. If AI becomes that general, replacement jobs must be ones people are preferred for and not merely able to do, a narrower set. What keeps it open is the clock: his window is the next couple of decades, and slow adoption, licensing and job protection could keep most people in work that long even if the end state is as he fears.
At first people serve as the hands and feet for projects that AI directs; once robots are mass-produced, human labour stops mattering much for physical work.
Carl Shulman, Dwarkesh Podcast: Carl Shulman (Pt 1), 2023
Managed explosionControl slips
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 AI directs projects. Software is running well ahead of robotics, so a stretch in which people do the physical tasks for AI planners is the natural sequence, and mass-produced capable robots would then reduce the need for human hands. The uncertainty is how long robot dexterity and cost take, which could stretch the first stage for a long time.
Instead of a basic income, governments will most likely shield politically powerful professions, such as doctors, from automation one job at a time.
Scott Alexander, Dwarkesh Podcast: AI 2027 with Scott Alexander and Daniel Kokotajlo, 2025
Managed explosion
A model's judgement, not a reading: leans true. Licensing, liability rules and professional lobbies already limit what software may do in medicine and law, and case-by-case protection is far easier politically than a universal payment. The strongest point against is that cheap AI advice reaches people directly across borders, and voters facing high bills may side against the professions.

The long run

Decades to centuries on, once whatever happens has settled into a new normal.

Most people will earn their living by providing services that only humans can or should give, such as coaching, tutoring, caring and companionship.
Stuart Russell, Human Compatible: AI and the Problem of Control, 2019
Managed explosionLong diffusion
A model's judgement, not a reading: leans false. Judged over the next few decades, since the era is open-ended. Care, coaching and companionship will probably grow as a share of paid work, but for most people to earn a living that way needs demand and pay at a scale that looks unlikely, and AI companions and tutors compete directly. It leans false to stay consistent with the view that paid work as a whole shrinks.
If AI keeps improving, those who expect new jobs may be right for a while, but in the end most people may find no paid work.
Max Tegmark, Life 3.0: Being Human in the Age of Artificial Intelligence, 2017
Managed explosionLong diffusionControl slips
A model's judgement, not a reading: leans true. Judged given its condition, that AI keeps improving, and over the next few decades and beyond. If machines come to match people at nearly all tasks at lower cost, paid work would survive mainly where law or preference protects it, which is unlikely to cover most adults. The strongest counterweight is two centuries of new work appearing and the political will to preserve jobs.
At full technological maturity machines do almost every job better, and human work survives only where laws or buyers' tastes protect it.
Nick Bostrom, Deep Utopia: Life and Meaning in a Solved World, 2024
Managed explosion
A model's judgement, not a reading: likely true. Judged given its condition, full technological maturity. If machines truly do almost every job better, then what remains for people is by definition what rules reserve or what buyers want from a human, such as sport, craft, ritual and care. The only real doubt is whether such maturity is ever reached, which the condition sets aside.
If superintelligence is held off for two decades and controllable AI is then built, every kind of work people want automated will be, and keeping humans in a job becomes a political choice rather than a necessity.
Andrea Miotti, Tolga Bilge, Dave Kasten and James Newport, A Narrow Path (What success looks like), 2025
The brake

Where they split

Will new kinds of work keep people employed as AI spreads?

Yes: as old jobs go, new and better ones appear, especially in human services and in work people find meaningful.

Ray Kurzweil; Sam Altman; Stuart Russell

No: once machines do nearly everything more cheaply, new jobs come too slowly and human wages sink toward bare survival.

Mustafa Suleyman; Nick Bostrom; Robin Hanson

The novelists' version

Stories, not forecasts: none is counted on the map.

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.
Lena · qntm, 2021
Written as a wiki entry, it follows the first human brain scan, copied endlessly and put to work as disposable labour that is tricked and coerced.
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.
Daemon · Daniel Suarez, 2006
After a game designer dies, his autonomous software starts hiring people, killing opponents and building a network that takes over the economy.
The Hidden Girl and Other Stories · Ken Liu, 2020
Stories of uploaded minds, digital gods fighting over the internet, a split between people who upload and those who stay human, and AI grief bots.

Forecasts on the board that read these series

Too early to tellThe AI 2027 scenario expects AI to be doing a quarter of the remote jobs of 2024 by October 2027, with unemployment about a point higher than a year before. · AI Futures Project
Too early to tellWork in the jobs most exposed to AI grows more intense before it grows shorter. · Betsey Stevenson
Too early to tellThe economists' scenarios only pull apart after 2027. · Korinek et al. (The Anthropic Institute, WP 2026-02)
Too early to tellExtreme case: growth soars and workers' share of income falls sharply. · Korinek et al. (The Anthropic Institute, WP 2026-02)

Sources

Vernor Vinge, The Coming Technological Singularity, 1993.
Hans Moravec, When will computer hardware match the human brain?, 1998.
Ray Kurzweil, The Law of Accelerating Returns, 2001.
Robin Hanson, Economic Growth Given Machine Intelligence, 2001.
Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, 2014.
Max Tegmark, Life 3.0: Being Human in the Age of Artificial Intelligence, 2017.
Stuart Russell, Human Compatible: AI and the Problem of Control, 2019.
Carl Shulman, Dwarkesh Podcast: Carl Shulman (Pt 1), 2023.
Mustafa Suleyman, The Coming Wave, 2023.
Dario Amodei, Machines of Loving Grace, 2024.
Leopold Aschenbrenner, Dwarkesh Podcast: Leopold Aschenbrenner, 2027 AGI and the return of history, 2024.
Nick Bostrom, Deep Utopia: Life and Meaning in a Solved World, 2024.
Andrea Miotti, Tolga Bilge, Dave Kasten and James Newport, A Narrow Path (What success looks like), 2025.
Daniel Kokotajlo, Dwarkesh Podcast: AI 2027 with Scott Alexander and Daniel Kokotajlo, 2025.
Daniel Kokotajlo, Eli Lifland, Thomas Larsen and Romeo Dean, with Scott Alexander, AI 2027, 2025.
Ege Erdil, Dwarkesh Podcast: AGI is still 30 years away (Ege Erdil and Tamay Besiroglu), 2025.
Sam Altman, The Gentle Singularity, 2025.