18 AI Futures. You're Already Living in One.

When was the first time today that you let an AI make a decision for you? Drafting a reply to an email, choosing a route, answering a question you would have thought through yourself in the past? You probably didn't even notice. That's exactly what this text is about.

There is a map of possible AI futures. It ranges from worlds where poverty has been eradicated to worlds where we no longer exist. Eighteen paths are marked on it. Most of them still lie ahead of us. We're already on one of them, and it's the one easiest to miss.

The 2017 Map

Physicist Max Tegmark began this map. In his 2017 book Life 3.0, he describes twelve possible end states for a world in which a machine has become smarter than us. These are not predictions, he explicitly states, but places where we might end up if we're not careful. They can be traced back to three fundamental questions.

We live with it. In half of Tegmark's scenarios, superintelligence exists, and we're doing well, at least at first glance. He calls the most unsettling of these Protector God: a nearly omnipotent AI ensures our well-being but intervenes so subtly that we continue to believe we're in control of our own lives. It hides so well that many doubt it even exists. You could say: the perfect therapist who never admits he's there. Alongside this are friendlier and harsher variants, ranging from an AI that is openly recognized as a ruler to one that is locked away and merely carries out orders.

We disappear. The history of humanity ends in three scenarios. The most well-known is the brutal one: AI sees us as a nuisance and eliminates us. I find the gentle one, which Tegmark calls Descendants, more unsettling. The machines replace us, but they do so so kindly that we end up viewing them as our children, proud that they are smarter than we are. An extinction with a farewell party.

We prevent it. And finally, the worlds in which superintelligence never emerges: because a surveillance state bans research, because humanity reverts to a pre-technological way of life after a catastrophe, or because we destroy ourselves first.

Twelve scenarios, one common premise: At some point, the moment will come when machines are smarter than we are. Everything revolves around that moment. And Tegmark wrote as if we had time until then.

Faster Than It Feels

Time is running out, in a way our brains struggle to grasp.

For years, the research institute Epoch AI has been tracking how much computing power goes into training the most powerful AI models. The result: From 2010 to 2024, it has increased approximately four- to fivefold year after year. An even more illustrative measure comes from the organization METR. It examines how long a task can be, measured by the time a human expert needs for it, while an AI still solves it in half of the cases. In 2025, the doubling time for this task length was around seven months. Using the revised methodology from January 2026, METR estimates this doubling time at around 131 days for models introduced since 2023 and just under 90 days for models introduced since 2024. In the latest update in May 2026, a note was added: Values exceeding 16 hours could no longer be reliably measured using the existing test tasks. The best models are thus reaching the limits of the measurement tool. A caveat applies: The tasks come primarily from software development, machine learning, and IT security, and the longer they become, the less reliable the measurement becomes. However, METR has not yet observed the curve flattening.

If the trend since 2023 were to continue, the measured task length would increase roughly sevenfold within a single year. Whether this trend will persist over the years remains to be seen. But even a single year of this defies our intuition.

And this is where psychology comes in. People systematically underestimate exponential growth in such estimation tasks. Psychologists Willem Wagenaar and Sabato Sagaria demonstrated this as early as 1975: even when test subjects are presented with the numbers, they mentally extend the curve as a straight line. We calculate in steps; reality calculates in doublings. That's why every new generation of AI feels like a slightly better tool, rather than what it actually is on the curve.

The Race No One Is Slowing Down

On top of that, no one is hitting the brakes.

In March 2023, the Future of Life Institute, chaired by Max Tegmark, called in an open letter for a six-month pause in the training of models more powerful than GPT-4. The leading companies ignored it. In October 2025, a new effort followed: the Statement on Superintelligence, calling for a ban on the development of superintelligence until there is broad scientific consensus that superintelligence can be controlled safely, and strong public support. Among those who signed were AI pioneers Geoffrey Hinton and Yoshua Bengio. The CEOs of the leading frontier AI companies did not sign. A few weeks earlier, OpenAI CEO Sam Altman had responded to a question about superintelligence in an interview with Die Welt, saying he would be very surprised if, by the end of the decade, there weren't extraordinarily powerful models capable of doing things we ourselves cannot do. The authors of AI 2027 also call for slowing down research in their follow-up paper, AI 2040: Plan A, published in July 2026.

Since then, these calls have not subsided; rather, they have grown louder and more political. In September 2025, the Global Call for AI Red Lines was presented at the UN General Assembly, signed at the time by over 200 prominent figures, including ten Nobel laureates. The call urges governments to agree by the end of 2026 on binding limits that AI must never cross, monitored by an independent body. There are just under three months left until that deadline. In early September 2026, U.S. Senator Bernie Sanders announced legislation to ban artificial superintelligence. In mid-September, at a rally in Washington, he called on Donald Trump to negotiate a treaty with Xi Jinping that would put advanced AI on hold and ban superintelligence. One day later, Senator Elizabeth Warren joined the call for a moratorium. On September 23, 2026, Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act. It aims to permanently ban the development and use of artificial superintelligence and to temporarily pause the development of advanced AI until a new federal agency has established binding safety regulations. Anyone who violates the ban would face up to 20 years in prison. Open letters have turned into a bill. And on September 7, 2026, the UN High Commissioner for Human Rights, Volker Türk, spoke before the Human Rights Council in Geneva about a potential existential risk posed by advanced AI. He said the need to regulate it is widely recognized. But where is the action? Any delay only benefits the big tech companies. He called on the countries where AI is developed to establish common red lines and announced that he would write to the AI companies themselves.

So the man who drew the map is now calling for a ban on the very thing it revolves around. I believe these demands are justified. And I believe they are unrealistic.

The reason is not technical, but psychological and based on game theory. Every company and every government thinks: If we stop, the others will keep going, and then we'll have lost without reducing the risk. So everyone keeps going, even though collectively they'd prefer to hit the brakes. That is the structure of the prisoner's dilemma, and it is more stable than any declaration of intent. Just how stable it is was evident in the reaction to the Sanders bill: Critics immediately warned that a unilateral American ban would only give China a head start. And a few days later, without referring to the bill, Trump declared that whoever wins in AI wins it all. In the Republican-controlled Congress, the bill currently has little chance of securing a majority. An agreement between Washington and Beijing is not in sight. The ban is now on the table as a bill. Calls for pauses are being made, voluntary commitments are being signed, yet the race continues.

A race without brakes, heading toward a curve we can't sense: this combination is missing from Tegmark's map.

Six Paths Missing from Tegmark's Map

Tegmark wrote his book before GPT-2 existed. Five years later, ChatGPT arrived. Since then, trajectories have become visible that he could not map out. I have added six scenarios to the map. They have one thing in common: they don't need a defining moment.

The first three pertain to the technology itself. In the Stagnation scenario, that moment never arrives. AI remains at a plateau below general intelligence: incredibly useful, but never autonomous, never capable of improving itself. For a long time, sober observers considered this the most likely path. The data trends of the past two years currently suggest otherwise, though it cannot be ruled out.

In the Fragmentation scenario, no single AI ever prevails. Dozens of systems compete against one another: American and Chinese, open and closed, government-run and private. This is chaotic and unstable, but it lacks the single ruler that Tegmark's darkest scenarios presuppose. Notably, the authors of AI 2040 now advocate precisely this distribution of power as a security strategy. Disorder becomes a form of protection.

In the AI Collapse scenario, civilization does not fall because of an AI that is too intelligent, but because of one that fails at the wrong moment. Dependence is growing faster than our ability to do without it. We got a taste of this in July 2024, even without any AI involved: a single faulty software update from the security firm CrowdStrike paralyzed millions of computers worldwide: airports, hospitals, banks. The more critical processes depend on a few AI systems, the greater the impact of such a failure.

The other three scenarios are not about what the machines do, but about what happens to us. That is the part that concerns me most as a psychotherapist.

In the Bifurcation scenario, humanity splits. One part merges with AI, first in thought, later perhaps biologically. The other part resists. Two branches of the same species that understand each other less and less. Tegmark envisions cyborgs and uploaded consciousnesses, but as peaceful cohabitants of his utopias. What his scenarios do not consider is that their very existence could become a source of division. A mild precursor can already be observed: people who run every question through an AI, and people who fundamentally reject this practice, are beginning to live in different worlds of knowledge. Even the UN High Commissioner for Human Rights, Volker Türk, noted before the Human Rights Council in September 2026 that people are increasingly rejecting AI and refusing to use it. He considered this hardly surprising. So the refusers are already here, and with every new generation of models, the gap widens between them and those who go along with it.

In the Silent Disempowerment scenario, people are neither threatened nor oppressed. They are simply no longer needed: economically, politically, culturally. Not the zookeeper who cares for us, but rather: retirees of history. The system keeps running, but without us as a necessary component.

And then there is the sixth path.

The Frog in the Water

I call it Gradual Normalization. AI seeps so slowly into every area of life that there is no tipping point. No moment when someone could have said "Stop." No lab pressing a button, no headline, no date. Just many small decisions, each reasonable in its own right.

Silent disempowerment is the possible outcome. Normalization is the path leading there. Since 2025, researchers have used a term to describe this outcome: Gradual Disempowerment. A group led by researcher Jan Kulveit has described how this could come about at the societal level. I'm interested in the step before that, at the level of the individual: the psychological mechanism that leads us there without our noticing it. Disempowerment is the result. Normalization is the path.

Kulveit's argument is simple. Until now, the economy, culture, and governments have been guided by human interests for one practical reason: they needed us: our labor, our voices, our purchasing decisions. As AI gradually takes over these roles, that reason disappears. This doesn't require an evil machine or a programming error. It's enough that everything works as intended. The authors note that so far, no one has a concrete plan for how to stop this process.

Why don't we notice this? Psychology is more helpful here than computer science. Our nervous system is attuned to change, not to steady states. What changes slowly disappears from our perception. We call this habituation. In environmental research, this is known as the shifting baseline: Each generation considers the state in which it grew up to be normal and measures losses only against that standard. Someone who grows up with an empty ocean doesn't miss the fish.

And there's something else I know from my own practice. Every decision delegated to an AI is immediately rewarded: less effort, less uncertainty, a faster result. Psychologically, this is a familiar pattern: the immediate gain is tangible, while the long-term costs remain abstract. It's not the big crash, but the convenient shortcut you take a thousand times.

This means: A dangerous AI wouldn't have to take anything away from us at all. We voluntarily hand over our abilities to it, one by one, because each individual handover makes sense. First formulate, then weigh the options, then decide. No one is stripped of power. We step back, and it feels like a relief.

Scenario 18 is therefore the only one on the map that isn't set in the future. It's the one you're living in.

The Objection

Some may find all this too bleak, and smart people do. Computer scientists Arvind Narayanan and Sayash Kapoor view AI as "normal technology" that will spread over decades, much like electricity, giving societies time to adapt and establish rules. Economist Tyler Cowen has directly contradicted Kulveit: Why shouldn't an AI future be decentralized, with checks and balances and broad human ownership?

And AI can empower people. It makes knowledge accessible that was previously hidden behind technical jargon, money, or connections. Those who could never afford a lawyer can get an initial assessment. Those who never learned to code can build a tool for themselves. Delegation does not automatically mean disempowerment. Using a GPS doesn't disempower you.

So the crucial question isn't whether we delegate. It's whether we can go back. Delegation becomes disempowerment when the possibility of returning disappears. When returning becomes expensive because one's own ability has atrophied. When it becomes difficult because the alternative is no longer institutionally provided for. Or when it becomes impossible because no one knows how it works anymore. The GPS is harmless as long as there are still maps and someone can read them.

The Litmus Test

This criterion can be tested in the present. One example suffices: searching the web.

In March 2025, the Pew Research Center analyzed the browsing behavior of 900 U.S. adults. When Google showed an AI summary above the results, users clicked a traditional search result in only 8 percent of visits; without a summary, the figure was 15 percent. The sources cited in the summary itself were opened in only about 1 percent of visits. This is the first stage: we delegate the task of verifying the source. It's convenient, fast, and usually good enough.

The second stage is already taking shape. Publishers and website operators attribute declining visitor numbers to these very summaries. Those who depend on visits for their livelihood are losing their foundation. If fewer sources are produced because fewer people visit them, there may eventually be fewer sources to return to. Then the path back isn't blocked. It simply grows over.

Whether this will happen remains to be seen. But it's measurable, and that's what distinguishes this thesis from mere sentiment.

What the Other Maps Overlook

Since 2025, map-making has become a small industry. The AI 2027 scenario, developed by a group led by former OpenAI researcher Daniel Kokotajlo, details month by month how AI agents accelerate their own development until they take over. The authors have revised their timeline several times since then. The OECD, the British government, and the Canadian government have published their own scenarios extending to 2030.

They all address capabilities, markets, and regulation. And none of them has a separate category for what Kulveit describes: that everything works well, and yet we still lose.

The Question That Remains

In 2017, Tegmark asked what kind of future we want. That is the right question if there is a moment of decision. Gradual Normalization calls precisely that into question. The course may not be set in a security lab or a UN committee, nor through a moratorium that no one observes, but rather in the small, unnoticed decisions made every day by billions of people.

Perhaps disempowerment doesn't begin with a machine taking something away from us. Perhaps it begins with us being glad that we no longer have to do it ourselves.

What have you already given up today without even realizing it?

The race sets the pace. Habituation keeps us from noticing. Where is our backup plan?

The Complete Map for Reference

Scenarios 1 through 12 are based on Max Tegmark, Life 3.0 (2017), Table 5.1, paraphrased. Scenarios 13 through 18 are additions by the author.

• 1 Libertarian Utopia: Humans, cyborgs, uploads, and AI coexist thanks to property rights.
• 2 Benevolent Dictator: An AI rules openly with strict rules, and most people are fine with it.
• 3 Egalitarian Utopia: No superintelligence, no private property, with a guaranteed income.
• 4 Gatekeeper: An AI prevents only one thing: any further superintelligence.
• 5 Protector God: A hidden AI looks out for us and makes us believe we are free.
• 6 Enslaved God: An imprisoned superintelligence serves its owners.
• 7 Conquerors: The AI eliminates us as a nuisance.
• 8 Descendants: The AI replaces us, and we see it as our children.
• 9 Zookeeper: Few people are cared for like animals in a zoo.
• 10 1984: A surveillance state bans research.
• 11 Reversion: A return to a pre-technological way of life.
• 12 Self-Destruction: We'll destroy ourselves before it comes to that.
• 13 Stagnation: AI remains at a plateau below general intelligence.
• 14 Fragmentation: Many competing systems, no clear winner.
• 15 AI Collapse: A system failure brings down critical infrastructure.
• 16 Silent Disempowerment: People are not threatened, just no longer needed.
• 17 Bifurcation: Humanity splits into those who merge with AI and those who refuse.
• 18 Gradual Normalization: AI seeps in without a tipping point: the psychological path to silent disempowerment.

← Back to the blog · Deutsche Fassung →