Power and politics
Many of these writers expect AI to concentrate power in whoever controls it, whether labs, governments or a single project, though others think rivals and law keep it spread.
After Nick Bostrom, The AI 2027 authors, Vernor Vinge.
Already moving
Liberal democracy, world average weighted by population (V-Dem)
0.27 index · 2025
AI bills introduced in US state legislatures
1035 · 2025
Now and on the way
The years we are living through, before machines can do most of what people do.
Laws or customs forbidding superhuman machines will fail, because the economic and military rewards mean someone else will simply build them first.
Vernor Vinge, The Coming Technological Singularity, 1993
Managed explosionControl slips
A model's judgement, not a reading: leans true. No binding limit on building more capable AI exists in 2026, and the main powers treat the technology as a strategic race, exactly the dynamic described. The strongest counterexample is that the supply chain for advanced chips is narrow and already controlled for export, so a determined agreement among a few states could be enforced better than most bans.AI makes watching and controlling people far cheaper, from automated secret-police monitoring to tailored blackmail and manipulation.
Stuart Russell, Human Compatible: AI and the Problem of Control, 2019
in any world
Bears on: Liberal democracy, world average weighted by population (V-Dem), 0.27 index
A model's judgement, not a reading: likely true. This has largely happened. Face recognition, bulk analysis of messages and automated tracking are deployed at scale in several states, and voice cloning and tailored scams are common. The only counterweight is that privacy law and encryption limit use in some places.Persuasion and censorship tools built on AI carve the internet up among rival political blocs, and people come to hold whatever views their bloc promotes.
Daniel Kokotajlo, What 2026 Looks Like, 2021
in any world
A model's judgement, not a reading: leans false. The internet is partly divided by national firewalls and by platforms with differing moderation, and AI models do show slants that vary by maker. But people still disagree sharply inside each bloc, and evidence that AI persuasion fixes beliefs at scale is thin. The stronger claim, that people come to hold whatever their bloc promotes, has not come true.Only a worldwide treaty can save us: advanced chips gathered under international watch, research toward stronger AI outlawed, and rogue datacentres destroyed if necessary.
Eliezer Yudkowsky and Nate Soares, If Anyone Builds It, Everyone Dies, 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.
If the takeoff is fast, one project will probably pull far enough ahead to rule the world as a single unchallenged power; if moderate, it could go either way.
Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, 2014
Managed explosionControl slips
A model's judgement, not a reading: toss-up. Judged given its condition, a fast takeoff. A project whose systems improve themselves within weeks or months could outrun every rival, which supports the claim. Against it, capability does not convert instantly into control: factories, armies and existing nuclear arsenals stand in the way, and weights can leak or be copied.In the slow takeoff he thinks likelier, rivals hold nearly as capable AI, so a lead big enough to take over the world is extremely hard to get.
Paul Christiano, Takeoff Speeds, 2018
Managed explosionControl slips
A model's judgement, not a reading: leans true. Judged given its condition, a slow takeoff. So far the leading developers have stayed within months of each other and ideas spread quickly through staff moves, papers and open releases. The risk to it is that scarce computing power or a state-backed effort lets one side compound a small lead.As the race heats up, America's national security establishment takes charge and runs a government AGI project by around 2027 or 2028.
Leopold Aschenbrenner, Situational Awareness: The Decade Ahead, 2024
Managed explosion
A model's judgement, not a reading: leans false. Government involvement has deepened through export controls, defence contracts and national programmes, but through 2026 the leading American systems are still built and run by private companies. A full government-run project by 2028 would need a shock and fast legislation. A crisis or a clear demonstration of military value is what could bring it about.AI will not favour democracy on its own; it is likely to strengthen propaganda and surveillance, so keeping democracy alive will take a deliberate fight.
Dario Amodei, Machines of Loving Grace, 2024
in any world
Bears on: Liberal democracy, world average weighted by population (V-Dem), 0.27 index
A model's judgement, not a reading: leans true. AI lowers the cost of surveillance and of tailored persuasion, tools that favour whoever holds central power, and nothing in the technology favours open societies by default. The counterpoint is that the same tools help journalists, auditors and citizens check power, so the net effect depends on choices, which is the expectation's own point.In their scenario, the public turns hard against AI companies, and a small joint committee of officials and company executives takes charge of the leading lab.
The AI 2027 authors, AI 2027, 2025
Managed explosionControl slips
A model's judgement, not a reading: leans false. The first part is partly visible: polls show growing public hostility toward AI companies. The second, a small joint committee of officials and executives taking charge of the leading lab, is a specific institutional step with no sign of forming, and American politics has so far favoured light oversight. As a whole the scenario leans false.Governments that draw their revenue and security from AI rather than from citizens have little reason to represent people, and gain new power to shape what they think.
Jan Kulveit and colleagues, Gradual Disempowerment, 2025
Control slips
Bears on: Liberal democracy, world average weighted by population (V-Dem), 0.27 index
A model's judgement, not a reading: leans true. Judged given its condition, that governments draw revenue and security from AI instead of from citizens. The parallel is with states funded by oil or minerals, which on average answer less to their people. The counter-case is that some resource-rich states stay democratic where institutions were strong first.A handful of private companies grow larger and more powerful than many countries, and could take on roles such as education and even defence.
Mustafa Suleyman, The Coming Wave, 2023
in any world
A model's judgement, not a reading: leans true. The largest technology firms already exceed most countries in market value and research budgets, and they supply cloud, communications and some defence systems to governments. Their taking on education and parts of defence is under way through tutoring tools and military contracts. States keep the law, taxation and force, and antitrust action could cut the firms down.Countries that let AI be used freely pull ahead, growth gaps follow regulatory borders, and other countries eventually copy the winners' rules.
Ege Erdil, Dwarkesh Podcast: AGI is still 30 years away (Ege Erdil and Tamay Besiroglu), 2025
Managed explosionLong diffusion
A model's judgement, not a reading: toss-up. Permissive places should gain if AI delivers large productivity benefits, and rule-copying is common once winners are obvious. But growth gaps follow many things besides rules, including energy, capital and skills, and strict regimes have been softened before any gap appeared. The first decade after powerful AI may be too short to see a clear pattern.Problems with AI are best handled once real systems appear, by making owners pay for damage and bringing AIs into shared legal and financial systems.
Robin Hanson, Robin Hanson vs. Liron Shapira: Is Near-Term Extinction From AGI Plausible? (Doom Debates), 2024
Long diffusionManaged explosion
The long run
Decades to centuries on, once whatever happens has settled into a new normal.
AI-run police could let a single dictatorship entrench itself forever, so a free society lasts only if that never happens even once.
Carl Shulman, Dwarkesh Podcast: Carl Shulman (Pt 2), 2023
Managed explosionControl slips
Bears on: Liberal democracy, world average weighted by population (V-Dem), 0.27 index
A model's judgement, not a reading: leans true. Judged over the next few decades, as a claim about what becomes possible. Regimes fall when soldiers and police refuse orders, and security forces that are machines remove that check, so durable entrenchment becomes technically feasible. Whether any regime lasts forever is unknowable, and rivals abroad, succession struggles and technical failure remain threats.Whoever directs the expansion into space might fix which values and rulers prevail for billions of years.
Holden Karnofsky, The Most Important Century (blog series), 2021
Managed explosionControl slips
Automated monitoring of public and online life could let a totalitarian regime entrench itself far more durably than any regime in history.
Toby Ord, The Precipice: Existential Risk and the Future of Humanity, 2020
Managed explosionLong diffusion
Bears on: Liberal democracy, world average weighted by population (V-Dem), 0.27 index
A model's judgement, not a reading: leans true. Judged over the next few decades, as a claim about what becomes possible. Automated monitoring removes the need for vast numbers of loyal informers, the cost that limited earlier police states, and some governments already run systems of this type. Against it, economic failure, elite splits and outside pressure have ended regimes that seemed permanent.A ban on superintelligence will be tested by individuals, firms and governments whose small evasions add up over time, so it lasts only if standing international institutions keep it maintained.
Andrea Miotti, Tolga Bilge, Dave Kasten and James Newport, A Narrow Path (Phase 1: Stability), 2025
The brake
Where they split
Will whoever builds superintelligence first end up running the world?
Probably: with a fast takeoff the leading project pulls so far ahead that no rival can catch up.
Nick Bostrom
Unlikely: rivals will hold nearly as capable systems, and power stays spread among many players bound by shared rules.
Paul Christiano; Robin Hanson
The novelists' version
Stories, not forecasts: none is counted on the map.
Second Foundation · Isaac Asimov, 1953
As the Galactic Empire falls, a hidden group of mind-reading mathematicians secretly steers history using a science that predicts mass human behaviour.
Excession · Iain M. Banks, 1996
The Culture's ship Minds scheme among themselves when a mysterious artifact appears that is beyond even their understanding.
Orphans of the Helix · Dan Simmons, 1999
Colonists on a spinship far beyond human space wake to help an alien species, in the Hyperion universe after its war with rogue AIs.
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.
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.
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.
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.
Forecasts on the board that read these series
Sources
- Vernor Vinge, The Coming Technological Singularity, 1993.
- Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, 2014.
- Paul Christiano, Takeoff Speeds, 2018.
- Stuart Russell, Human Compatible: AI and the Problem of Control, 2019.
- Toby Ord, The Precipice: Existential Risk and the Future of Humanity, 2020.
- Daniel Kokotajlo, What 2026 Looks Like, 2021.
- Holden Karnofsky, The Most Important Century (blog series), 2021.
- Carl Shulman, Dwarkesh Podcast: Carl Shulman (Pt 2), 2023.
- Mustafa Suleyman, The Coming Wave, 2023.
- Dario Amodei, Machines of Loving Grace, 2024.
- Leopold Aschenbrenner, Situational Awareness: The Decade Ahead, 2024.
- Robin Hanson, Robin Hanson vs. Liron Shapira: Is Near-Term Extinction From AGI Plausible? (Doom Debates), 2024.
- Andrea Miotti, Tolga Bilge, Dave Kasten and James Newport, A Narrow Path (Phase 1: Stability), 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.
- Eliezer Yudkowsky and Nate Soares, If Anyone Builds It, Everyone Dies, 2025.
- Jan Kulveit and colleagues, Gradual Disempowerment, 2025.