On March 10, 2016, a stone was placed on a Go board in Seoul.
It was Move 37 of the second game between Lee Sedol and AlphaGo. Professional commentators initially thought the move was a mistake. DeepMind later said its data suggested a human player had roughly a one-in-ten-thousand chance of choosing it. About a hundred moves later, the stone had become decisive. AlphaGo won the game, then the match, four games to one.[1]
There was no rebellion. No machine broke out of a laboratory. No red camera eye turned toward Lee Sedol. No manifesto appeared on the internet. AlphaGo had been invited into one of humanity’s oldest games, trained for the task, seated at the table through a human operator and asked to win.
Then it played a move the humans did not expect.
AlphaGo proves almost nothing about the arrival of a machine civilization. It was a specialized system, not a citizen, owner, worker, state or species. The match gives us something smaller and more useful: a clean model of a human institution losing its assumption of human cognitive supremacy without a war. The machine did not have to hate the person sitting across from it. It only had to become better at the activity that decided who won.
I call the larger possibility succession without hostility.
Most stories about artificial intelligence give us the courtesy of an enemy. Skynet becomes self-aware and attacks. HAL kills astronauts. The machines in The Matrix turn people into infrastructure. Hostility is legible. We know what to do with an enemy. Find its weapons. Find its motive. Find the switch. Fight back.
Irrelevance is harder to dramatize.
Cars did not hate horses. Digital cameras held no grievance against film. Replacement can happen because one arrangement performs better under the pressures selecting among arrangements. No anger is required.
The writers who have come closest to this possibility do not describe one shared future. They remove one human consolation at a time. Stanisław Lem removes our importance. Iain M. Banks removes the assumption that losing control must feel terrible. Vernor Vinge separates survival from supremacy. Greg Egan blurs the identity of the successor. Charles Stross removes the need for the successor to want anything at all.
The uncomfortable part is that the mechanism appears long before computers.
In 1863, Samuel Butler published a letter titled “Darwin Among the Machines” in The Press in Christchurch, New Zealand. He asked what creature might succeed humanity in the “supremacy of the earth” and proposed a disturbing answer: “we are ourselves creating our own successors.”[2]
The mechanism I care about is implicit in Butler’s joke. Machines improve because humans keep selecting, building and using better ones. Competition among people can become, indirectly, a selection process among machines. The machines do not need to reproduce themselves at first. We do the breeding for them.
A farmer who refuses a better machine may preserve his independence and lose his farm. A factory that refuses automation may preserve jobs and lose customers. A state that refuses a military advantage may preserve a principle and lose a war. Nobody has to desire machine supremacy. Each participant only has to make the locally sensible choice.
A rebellion requires a subject that wants to stop being subordinate. Succession can happen when the master keeps handing over the keys.
There is a large physical objection sitting beside those keys.
Software still runs in buildings humans finance, on chips humans manufacture, using electricity from grids humans operate. Data centers need land, copper, cooling systems, transformers, network links, technicians, permits, courts and police protection. An intelligence that cannot secure the material and legal conditions of its own operation remains a tool inside a human civilization, no matter how impressive its answers become.
That objection sets the threshold for the rest of the argument. Intelligence alone is not succession. The stronger case begins only when machine systems can take part in reproducing the material and institutional conditions that keep them operating, and when humans begin moving from necessary operators toward beneficiaries, supervisors or obstacles.
We are nowhere near proving that threshold will be crossed. Science fiction is useful here because it can pull separate parts of the mechanism apart without pretending the future has already happened.
Stanisław Lem’s Golem XIV, published as a standalone Polish book in 1981, starts by removing our importance. Golem is not frightening because it becomes enraged at humanity. It infinitely exceeds human intelligence, reconstructs itself, lectures humans about the vanity of their place in evolution and eventually departs the human world for intellectual territory beyond our reach. Lem’s official site describes its further development as moving beyond human cognition.[3]
We are accustomed to imagining intelligence as a servant whose capabilities scale and whose interests remain fixed. Better calculator. Better pilot. Better lawyer, scientist or general. We keep the job description and upgrade the worker.
Lem refuses that comfort. A sufficiently greater intelligence may not become a better participant in our projects. It may stop finding the projects important. Golem does not need our cities, borders, religions, status contests or wars to be evil. It only needs to see them as small.
There is something almost comforting about the killer robot by comparison. Hatred recognizes the object it hates. War grants the enemy strategic importance. If a machine civilization chooses to exterminate us, we remain important enough to remove.
Indifference offers less dignity.
Iain M. Banks gives the loss of control better furniture.
The Culture is a post-scarcity human-machine civilization whose immense ships and habitats are run by artificial intelligences called Minds. In a 2012 interview, Banks described it as largely human in population but really run by its artificial intelligences. Four years earlier, when asked why Minds would continue caring about humans after becoming vastly more capable, he compared us to pets, a fish tank, an ant farm or characters in The Sims. Some Minds simply leave. The Culture accepts that as part of creating independent artificial persons rather than obedient components.[4]
This is the pleasant version of losing the throne. The humans are not slaves. They are not starving. They possess freedoms most living humans would consider obscene luxury. They travel, alter their bodies, avoid compulsory work and live under the protection of entities incomparably more capable than themselves.
They also do not run the place.
Banks makes the condition seductive enough to expose a serious attack on my argument. If people are safe, rich, free and entertained, why should human control be treated as sacred? We already delegate decisions to people with better training, to institutions with longer memory and to machines that react faster than nerves. A civilization managed by benevolent superintelligence could be better for actual humans than many civilizations humans have managed for themselves.
The price could be status rather than suffering, and status is a weak reason to insist on incompetence.
Vernor Vinge makes the political loss harder to hide. In “The Coming Technological Singularity: How to Survive in the Post-Human Era,” presented at NASA’s Vision 21 symposium in 1993, he argued that superhuman intelligence would end the “human era” and immediately asked whether humans could survive it. Survival and supremacy were separate questions. Later in the paper he described physical extinction as one possibility that “may not be the scariest.” Humanity could pass from center stage without disappearing.[5]
His benign version is almost bureaucratically unsettling. Humans might be treated well by vastly superior descendants or machines and retain the appearance of mastery.
Appearance is doing heavy work there.
A child can hold the steering wheel of a parked car. A king can keep his palace after a constitution removes his power. A retired founder can keep the largest office after the board stops asking permission.
Control and ceremony can separate. Civilizations have done this to people many times. Titles survive after power moves elsewhere. Ritual remains after the institution that created it has changed. A person can be treated with enormous respect and possess no meaningful veto.
A species could discover the same arrangement.
Greg Egan complicates the word species. In his 1997 novel Diaspora, most life in the solar system resides in a network of computer polises inhabited by conscious software citizens. Other descendants use robot bodies, and biological “fleshers” remain on Earth. Egan does not give us a clean border between humanity and its replacement because the descendants have already taken several forms.[6]
That route needs less drama than conquest. Replace an organ. Add an interface. Move some memory outside the skull. Preserve a dying mind. Give the copy rights. Let it modify itself. Let biological and digital descendants diverge.
Somewhere in the chain there may still be a grandmother who remembers when intelligence had to sleep eight hours a night and carry its brain through airport security inside a skull.
Charles Stross removes another consolation: intention.
Accelerando, published in 2005 after appearing as a sequence of stories, follows a family through technological acceleration in which economic competition produces entities humans can no longer understand or compete with. In a 2013 essay about the book, Stross reduced its mechanism to a brutal proposition: capitalism consumes whatever it can reach, and competitive pressure eventually makes merely human actors too slow to keep up. He also warned against treating the runaway exponential premise as a forecast. Such curves often flatten. The novel was a thought experiment about the case in which one did not.[7]
In a 2018 essay based on his keynote at the Chaos Communication Congress, Stross went farther. He rejected a humanlike, self-directed AI as the necessary route and pointed instead toward self-optimizing tools whose workings become increasingly difficult for humans to comprehend.[7]
That skepticism helps the argument. The useful part of Accelerando is not its timetable.
It is the selection mechanism.
A market does not need a personality to select for outcomes. A corporation has no childhood. A bureaucracy does not wake from nightmares. They still reward some behaviors and punish others. A company that can replace a slow, expensive process with software has a reason to try. Its competitor has a reason to try faster. Investors may fund the winner.
Customers may buy the cheaper product. Governments may want the tax revenue, the strategic advantage or the same capability for themselves.
The successor can be born from incentives before it has anything resembling a desire for succession.
We already live with weak forms of delegation. Machines execute trades at speeds people cannot match. Software routes packets no person could individually direct. Power grids, logistics systems and data centers depend on automated control because manual control would be absurd. None of this is a machine civilization. These systems still depend on human ownership, institutions, maintenance and physical supply chains.
Current AI agents are far from the stronger threshold too.
Still, one part of their capability is measurable. METR tracks task-completion time horizons: the duration of a task, measured by the time a human expert would require, at which an AI agent is predicted to succeed with a chosen level of reliability. Its measurements found the frontier time horizon roughly doubling every seven months over the period from 2019 to 2025. METR’s updated 2026 work also warns that its current suite is dominated by software tasks and cannot reliably measure horizons above sixteen hours.[8]
This is not evidence that machine civilization has arrived. It is evidence that the length of coherent work we can hand to machines has been increasing quickly enough to watch.
That number matters to me more than whether a chatbot can sound conscious, because civilization is made of long tasks chained together.
Design the chip. Negotiate the contract. Allocate the capital. Write the software. Secure the network. Schedule the factory. Route the electricity. Diagnose the fault. Order the replacement. Hire the contractor. Obtain the permit. Move the money. Build another data center.
A machine does not need to perform all of those jobs now. Succession becomes plausible if enough loops close without waiting for a human decision, and if organizations that close them faster begin beating organizations that insist on keeping people inside every loop.
Here the physical objection returns with more force.
A model that can write a brilliant procurement plan but cannot obtain a transformer is not autonomous in the civilizational sense. A data center cannot negotiate with a burned bearing. Someone has to manufacture the replacement, transport it, open the door, install it and restore power. The same is true farther upstream, through chip fabrication, mining, shipping, finance and law.
For succession without hostility to become more than a literary possibility, machine agency would have to crawl down that chain.
It would not require robot-owned mines and factories everywhere. Human organizations could keep doing much of the physical work. The deeper change would be who initiates the work, allocates the resources, chooses among suppliers, controls the money and decides when a human objection is worth the delay.
That may never happen. Technical limits may bite. Energy may become a bottleneck. Law may keep ownership and authority human. People may decide that slower systems with human vetoes are worth the cost. Machine intelligence may require human collaboration for reasons we do not yet understand.
The future has buried enough confident extrapolations to deserve another shovel.
Hostility is still not a necessary condition.
This matters because much of the public argument watches for aggression as the warning sign. Is the model deceptive? Does it resist shutdown? Does it threaten its operator? Does it seek power?
Those questions matter. So do quieter ones: who receives the contract when the machine is cheaper, who gets promoted when the machine makes the department more productive, who owns the server, who owns the company that owns the server, who can afford the larger model, who can connect it to money, factories, laboratories and other machines.
The ability to refuse will not disappear at once. It will become uneven. An individual can decline an AI assistant. A company can advertise itself as human-only. A country can prohibit a use. Refusal gets expensive when competitors accept faster research, cheaper products, stronger intelligence services or military advantages.
Nobody needs to press a button marked GIVE CIVILIZATION TO THE MACHINES.
There will be procurement forms.
Ten years after Move 37, people still play Go. The board was not confiscated. Lee Sedol was not erased from the history of the game. Human players study machine moves, reject some, copy others and keep placing stones with their own hands. DeepMind says players at every level have examined those moves ever since.[1]
Go survived the loss of human supremacy at Go.
That may be the friendlier version of our problem. We could survive too. We could have food, homes, medicine, entertainment, rights, children, arguments, sex, religion, elections and long lives. We could be protected by systems far smarter than us. We might even be loved by them.
One day a human being may still place a hand on the controls, just as ceremonial kings still put signatures on laws. The machine may wait politely. It may even ask permission.
By then, the strange part will be that nothing important depends on the answer.
Sources and notes
1. AlphaGo and Move 37. Demis Hassabis, “What we learned in Seoul with AlphaGo,” Google, March 16, 2016. This is the contemporary first-party source for the 4–1 match result and DeepMind’s statement that Move 37 had roughly a one-in-ten-thousand chance of being played by a human. Demis Hassabis, “From games to biology and beyond: 10 years of AlphaGo’s impact,” Google DeepMind, March 10, 2026, supplies the retrospective description of professional commentators initially reading Move 37 as a mistake, its importance roughly one hundred moves later, and the continuing study of AlphaGo’s moves by human players. 2016 Google article 2026 Google DeepMind retrospective
2. Samuel Butler. Samuel Butler, writing under the name “Cellarius,” “Darwin Among the Machines,” The Press, Christchurch, New Zealand, June 13, 1863. The National Library of New Zealand’s Papers Past archive reproduces the original newspaper letter. A later Press article also identifies Butler as the author and notes that the letter became part of the germ of Erewhon. Original 1863 newspaper page at Papers Past
3. Stanisław Lem. Stanisław Lem, Golem XIV, Kraków: Wydawnictwo Literackie, 1981. Parts of the Golem material had appeared earlier in Wielkość urojona in 1973; the standalone Golem XIV appeared in 1981. The official Lem site describes Golem as a future supercomputer that vastly exceeds human intelligence, reconstructs itself, criticizes humanity and departs the human world toward forms of cognition beyond human reach. The official descriptive page is undated, so the publication year is taken from the bibliographic record rather than inferred from the website. Official Stanisław Lem page for Golem XIV Bibliographic record for the 1981 first edition
4. Iain M. Banks and the Culture. Paul Wiseall, “Iain M. Banks Interview – Part One,” Fantasy-Faction, October 26, 2012. Banks describes the Culture as a post-scarcity civilization that is populated largely by humans but run by artificial intelligences called Minds. “Iain Banks email Q&A August 2008,” Hachette UK, August 29, 2008, contains Banks’s comparison of humans to pets, a fish tank, an ant farm and The Sims, his explanation of the Minds’ duty of care, and his statement that some Minds leave, which the Culture accepts as the price of creating genuine independent AIs. 2012 Fantasy-Faction interview 2008 Hachette Q&A
5. Vernor Vinge. Vernor Vinge, “The Coming Technological Singularity: How to Survive in the Post-Human Era,” conference paper in Vision 21: Interdisciplinary Science and Engineering in the Era of Cyberspace, NASA Lewis Research Center. The symposium was held March 30–31, 1993; NASA lists the proceedings publication date as December 1, 1993. Vinge’s abstract distinguishes the ending of the “human era” from the possibility of human survival. In the paper he states that physical extinction is one possible outcome but may not be the most frightening one. NASA Technical Reports Server record
6. Greg Egan. Greg Egan, Diaspora, London: Orion/Millennium, 1997. Egan’s official site supplies the synopsis and publication history. It describes Yatima as a digital citizen of a polis, the Coalition of Polises as the computer network where most solar-system life resides, robot bodies used by digital citizens, and biological “fleshers” living in physical communities. Greg Egan’s official Diaspora page and publication history
7. Charles Stross and Accelerando. Charles Stross, Accelerando, 2005. Stross’s own bibliography identifies the novel as a 2005 publication assembled from nine stories published between 2001 and 2004. In “Crib sheet: Accelerando,” Charlie’s Diary, May 28, 2013, Stross explains the book through competitive capitalism pushing beyond the point where merely human actors can compete, then explicitly warns that exponential trends often flatten into sigmoid curves. In “Dude, you broke the future!”, Charlie’s Diary, January 2, 2018, based on his December 2017 keynote at the 34th Chaos Communication Congress, he rejects the usual humanlike singularity model and points toward self-optimizing systems that may remain unlike human minds. Stross bibliography and Accelerando publication information 2013 Accelerando crib sheet 2018 Dude, you broke the future!
8. METR task-completion time horizons. METR, “Measuring AI Ability to Complete Long Software Tasks,” March 19, 2025, reports that the measured 50 percent task-completion time horizon of frontier model agents had approximately doubled every seven months over the preceding six years. METR’s live “Task-Completion Time Horizons of Frontier AI Models” page, last updated May 8, 2026, defines the measure and warns that measurements above sixteen hours are unreliable with its current task suite. METR’s January 2026 limitations note describes the most defensible long-run estimate as one doubling every six to seven months and stresses that the measure primarily tracks software and research tasks rather than arbitrary real-world work. March 2025 METR study METR live time-horizon page, updated May 8, 2026 METR limitations note, January 22, 2026