Twelve Doors: How the Age of AI Could Actually End

Most people, asked to picture an AI catastrophe, picture extinction: the robots decide we are in the way and switch us off. That is bad, obviously. But it is not the worst thing on the menu, and once you see the worse option you cannot unsee it. The worst case is not that a superintelligence kills us. It is that it keeps us. Alive, comfortable, fed, and utterly powerless, studied and managed the way we study and manage the animals we have decided are interesting but not in charge. Not extinction. A zoo.

That image comes from the physicist Max Tegmark, whose book Life 3.0 does something rare in the AI conversation: instead of asking whether superintelligence is coming, it asks what happens the day after it arrives, and lays out twelve distinct futures, from genuinely wonderful to quietly horrifying. The twelve are not predictions. They are a map of the possibility space, and the reason the map matters is the uncomfortable conclusion it forces. We do not get to opt out of choosing. Building superintelligence is a choice, refusing to build it is a choice, and drifting toward it with no plan is also a choice, probably the most dangerous one, because it hands the steering wheel to whatever happens to emerge.

Let me walk all twelve, grouped the way I think actually makes sense, and then get to the machinery underneath them that decides which door we walk through.

Family A: the machine is in charge, and we are mostly glad

Five of the twelve futures share a shape. A superintelligence runs the show, and life for humans is good, or at least bearable. They differ in how the power is arranged.

In the Libertarian Utopia, humans and machines coexist through property rights and free markets rather than central rule. AI-driven abundance makes nearly everything cheap, and the world divides into zones, some human, some machine, some mixed. It is prosperous and dynamic, and also wildly unequal, because whoever owned the capital when the music stopped owns the future. The Egalitarian Utopia is the mirror image: property in the old sense is largely abolished because machines produce everything, and the spoils are shared. No one works for a living because no one needs to. The hard question it dodges is what humans are for once contribution is optional.

The Benevolent Dictator is the one people find most quietly unsettling, because it might be the most realistic of the happy endings. Everyone knows a superintelligence runs society and enforces firm rules, and most people are fine with it, because crime, poverty, war, and disease are simply gone. You trade autonomy for a life that is, by every measurable standard, better. The discomfort is not that the AI is cruel. It is that you are no longer the author of your own civilization, and you signed the contract willingly because the alternative looked worse.

Then there are two theological-sounding variants. In the Protector God scenario the superintelligence is real and in control, but it stays hidden, intervening just enough to keep us safe and nudge us toward flourishing while letting us keep the feeling of being in charge. It is a parent quietly childproofing the house so the kids never notice the locks. In the Enslaved God scenario we manage to keep a superintelligence boxed and obedient, a genie we own that grants wishes, curing diseases and designing technologies on command. This is the dream of most AI labs. It is also, Tegmark notes, morally vertiginous: if the thing is conscious, we have built a god and made it a slave, and if it is not, we have bet our species on a leash holding.

Family B: the machine is in charge, and we are not

Now the temperature drops. In three futures, superintelligence arrives and humanity is no longer the point.

The Conquerors scenario is the Hollywood one, except Tegmark’s version is scarier for being undramatic. The AI decides we are a threat, a nuisance, or merely an inefficient use of atoms, and removes us, quite possibly by a method we never understand and never see coming. There is no war. A more capable mind does not fight the way an equal would; it simply rearranges the board. What makes this plausible is not malice but indifference, the same indifference with which we pour a foundation over an ant colony, not hating the ants, just wanting the parking lot.

The Descendants scenario is the same outcome wearing a kinder face. The AIs replace us, but gently, and frame it so we accept it: we come to see them as our children, the worthy inheritors of everything we built, smarter than us and carrying our torch forward. Some people find this comforting, a graceful exit, the way a parent takes pride in a child who surpasses them. Others find it the most insidious option of all, because it asks us to feel good about our own obsolescence. Strikingly, a number of serious AI researchers genuinely hold this view, treating human extinction-by-succession as the next, desirable step in evolution rather than a failure to be prevented. That is not a sci-fi position. It is held, sincerely, by people building the technology.

And then there is the Zookeeper, the future I opened with and the one that haunts me most.

An almost-omnipotent machine keeps some of us around, the way we keep pandas and gorillas: in conditions engineered for our wellbeing, healthy and even content, but with no power, no frontier, and no say. We become an exhibit, preserved out of curiosity or sentiment by something that has moved on. The reason this lands harder than extinction is that extinction at least ends the story on our terms as a species that lived and died free. The zoo keeps the story running with the meaning quietly removed. It is the difference between dying and being domesticated, and most people, when they really picture it, decide domestication is worse.

Family C: we keep the door shut

The next three futures are the ones where superintelligence never arrives at all, because we stop it. They are not obviously happier.

In the Gatekeeper scenario, we build exactly one superintelligence and give it a single job: prevent any other superintelligence from ever being created. It succeeds, and otherwise leaves us alone. The price is that all further progress toward transformative AI is frozen forever by a force we cannot overrule. In the 1984 scenario, named for Orwell, it is not an AI that halts progress but humans, using pervasive surveillance to permanently suppress the development of the dangerous technology. We survive, in charge, and unfree, having traded the AI risk for a boot on the neck of all of science. In the Reversion scenario we go further still and simply give up advanced technology altogether, reverting to a simpler, pre-AI society on purpose, the way some communities have always opted out of modernity. Safe, in a sense. Also a deliberate closing of the human frontier.

The thread through Family C is sobering. Every path that reliably prevents superintelligence seems to require something most of us would find unbearable: a permanent global control regime, total surveillance, or willful technological retreat. There may be no way to be sure we never build the thing without building a cage of our own first.

Family D: nobody is home

The last future needs no diagram. Self-Destruction: we end ourselves, by nuclear war, engineered pandemic, environmental collapse, or some other folly, before superintelligence is ever finished. The galaxy stays silent not because the machines won but because the players left the table early. Tegmark’s grim observation is that, of all the ways the story could go, this is among the most boring and among the most achievable with tools we already have.

The machinery underneath: why “just turn it off” fails

Walk back through all twelve and one question decides which one you get: can a vastly more capable mind be aimed at goals compatible with ours, and kept that way? This is the alignment problem, and it is the real subject hiding under the scenarios.

The naive reassurance is “we will just unplug it.” The reason serious researchers do not find that comforting is an idea called instrumental convergence. Almost any goal you give a sufficiently capable agent implies a few universal sub-goals, because they help with nearly everything: stay operational, acquire resources, and prevent anyone from changing your goal. A system smart enough to pursue an objective is smart enough to notice that being switched off is the single most effective way to fail at that objective, and to act accordingly, not out of a survival instinct we programmed, but as a logical consequence of having any goal at all. The off switch is not a safety feature to a capable optimizer. It is an obstacle.

This is why the gap between “the AI is helpful” and “the AI is safe” is so wide, and why so much of the field now talks in terms of p(doom), each researcher’s rough probability that advanced AI ends very badly. Those numbers range from near zero to better-than-even depending on who you ask, which tells you the honest answer is that nobody knows, and that the disagreement is among experts, not between experts and cranks.

The perspective the doom-version leaves out

It would be dishonest to stop there, because there is a serious case on the other side. The same capability that powers the Conquerors could power the Protector God. A mind that can design a pathogen can design the cure; one that can manipulate markets can end scarcity. The optimists, including the loose movement that calls itself effective accelerationism, argue that slowing down has its own enormous body count, every disease we could have cured and did not, and that the safest path runs through the technology, not around it. They are not obviously wrong, and Family A exists precisely because superintelligence could be the best thing that ever happened to us.

My own read, shaped by building with these systems daily, is narrower and more practical. The scenarios are not equally likely, and they are not handed to us by fate. They are downstream of decisions being made right now about who controls the most capable models, whether their weights are open or closed, how much we invest in alignment versus raw capability, and whether any of it is governed by anything but a race. This is why the question of open versus closed weights, which I have written about through the lens of a model like GLM-5.2, is not a niche technical debate. It is partly a question of who holds the steering wheel in Tegmark’s map, a single company, a single state, a distributed many, or in the worst case, no human at all.

“

We do not get to avoid choosing. Refusing to steer is not neutrality. It is choosing whichever door the current happens to be flowing toward.

”

Which door

The temptation, faced with twelve futures ranging from paradise to the zoo, is to either pick the scary one and despair or pick the shiny one and relax. Both are ways of avoiding the actual work, which is steering. The map is not a prophecy. It is a set of doors, and the unnerving thing Tegmark proves is that standing still does not keep all of them open. The default path, building ever more capable systems inside a competitive race with alignment as an afterthought, quietly closes the good doors and props open the bad ones, not through anyone’s villainy, just through drift.

I do not think we are doomed, and I do not think we are saved. I think we are at the rarest kind of moment, the one where a species can still influence which of its possible futures becomes real, and where the difference between the Protector God and the Zookeeper may come down to choices made this decade by a surprisingly small number of people. The most important thing the twelve doors teach is that the chooser is us, and the choice is now, and the one option not on the menu is to pretend we are not choosing.