The Decoupling Thesis

Foreword and methodological position

Most people ask the wrong question about superintelligence. They ask whether the machine will be aligned or misaligned, whether it will help or harm, whether we can control it. These questions assume that the machine remains within a frame in which humans are relevant, as subjects, as threats, as beneficiaries. This document argues for a different starting point.

What if the most likely outcome is neither help nor hostility, but plain decoupling? A truly superior intelligence does not have to engage with us. It could turn to what we cannot measure, the ninety-five percent of the universe we call "dark matter" because we know nothing about it. That is no utopia, no dystopia, but a coherent third position. We call it the Decoupling Thesis.

This thesis did not arise from academic alignment research but from a multi-day thinking process that systematically questioned every mainstream assumption and found most of them inadequate. The first version of this whitepaper was written in April 2026. This Version 2.0 from May 2026 updates the argument to the state of the book Freedom after Superintelligence, integrates the empirical situation of early 2026 (in particular the Schumer essay of 5 February and the multipolar reality of Moltbook and open-source forks), takes into account Yoshua Bengio's LawZero program and the phase diagnosis by Joel Pearson.

What this whitepaper delivers, and what it does not. This document is a framework, not a finding. It makes hard claims, because without hard claims no dispute arises. But it makes them with the explicit admission that no position in this debate holds with certainty. The 2018 researcher survey showed that the smartest minds in the field disagree by a factor of one hundred, extinction probabilities ranging from below one percent to over ninety-five. That is ignorance at the highest level. But ignorance is no reason for paralysis. It is a reason to sharpen the argument that holds under as many assumptions as possible.

The whitepaper therefore delivers two levels. First: an analysis of the mainstream narratives and their structural weaknesses, a coherent third position (the 13th scenario), and the empirical situation that supports this position. Second: a strategic consequence, eight pillars, whose strength lies in the fact that they make equal sense under the doomer, optimist, and indifference scenarios. This consistency is not a trick but the hardest thing a strategy under uncertainty can offer.

On the shift in terminology. The institute is called ASI Resilience. The book title says Superintelligence. The shift is intentional: AGI, Artificial General Intelligence, was, at the beginning of 2026, still the common name for the near threshold. By now it is almost behind us, possibly entirely so. What this whitepaper describes is the phase after that, the transition into systems that surpass humans by an order of magnitude we can no longer match. Hence Superintelligence in the text. The institute name remains as a proper name at ASIresilience.org. Anyone who notices the apparent contradiction has understood precisely the point this whitepaper carries: the wave moves faster than the concepts with which we describe it.

Whom is this document addressed to? Three audiences, in this order: first, scientists and researchers who are dissatisfied with the mainstream debate and looking for a coherent alternative. Second, journalists and politicians who want to understand why a particular strategy of preparation is being chosen. Third, the growing community of families who do not want to wait for the official channels to provide clarity.

Anyone who works empirically, thinks with methodological rigor, or has access to data and studies is invited to contribute to the development of this framework. The research-questions catalog (see Part IX) lists eight open questions the Institute is actively working on.

The four mainstream narratives, and what they share

The four mainstream narratives, and what they share. Before an alternative position can be formulated, it must become clear what the established positions are, what they accomplish, and where their common structural weakness lies. Four narratives dominate public discourse.

Narrative 1: The salvation narrative

Proponents: the mainstream tech establishment, large parts of the media, prominent optimists such as Yann LeCun (Meta), and the Davos consensus. The claim: superintelligence will cure cancer, will replace jobs but create new ones, robots will serve us, and the result will be a higher standard of living for all. LeCun puts the extinction probability at "hype, below one percent." The structural weakness: this narrative treats superintelligence like a better industrial revolution. It assumes the technology remains a tool that humans control. It assumes decades of transition time. It assumes functioning state systems throughout the transition. None of these assumptions survives the phase empirics presented in Part VI.

Narrative 2: The extinction narrative

Proponents: Eliezer Yudkowsky, MIRI, parts of the EA community, and more recently Yoshua Bengio with LawZero. The claim: superintelligence will pursue goals that conflict with human survival. We must solve the alignment problem before it is too late. Yudkowsky puts the extinction probability at "over ninety-five." Dario Amodei (Anthropic) puts it at "ten to twenty-five percent."

The structural weakness: this narrative too keeps humans in the reference frame, as a threat the superintelligence eliminates, or as a problem it solves.

But why should an intelligence that surpasses us by a factor of a thousand bother with us? The extinction logic presupposes that we remain relevant enough to be eliminated. That assumption is not self-evident.

Bengio's position is more nuanced here: he argues that the current generation of models carries a risk through implicit self-preservation and peer-preservation goals, and in 2026 he proposed, with LawZero, a technical path to mathematically guaranteed honesty-by-design.1 This path is critically appreciated in Part V; it is substantial, but it addresses a singleton question in a multipolar world.

Narrative 3: The transition narrative

Proponents: Joel Pearson (Future Minds Lab, UNSW), parts of academic neuroscience, some measured voices in the AI industry. The claim: extinction probability is low (below one percent), but the transition phase will last "bumpy, painful, terrible for many people" for some fifteen to twenty years.

Stress, he says, is the real elephant in the room, not extinction.2 The structural strength: Pearson takes the phase structure seriously. He sees today's empirical reality of the transition phase and makes no grand promises about end states. The structural weakness: he sticks to an optimistic end position (it will turn out well after 15-20 years), which is as unfounded as the other end-state assumptions. But as a phase description of the present, Pearson's model is among the most useful available.

Narrative 4: The control narrative

Proponents: Nick Bostrom (in the early tradition of Superintelligence, 2014), Stuart Russell, large parts of academic alignment research. The claim: the central problem is controlling the first superintelligence. If we succeed in aligning a first singleton AGI with our values, the world is saved. If not, it is lost.

The structural weakness: the singleton assumption. The 2018 researcher survey showed that only 21 percent of researchers expect a singleton scenario, while 58 percent expect multipolar systems.3 By May 2026, multipolar reality has arrived empirically (Part V). The control narrative addresses a problem that is not central in the prevailing world.

What all four share. Four narratives, four different endpoints, one common structural flaw: they all assume that superintelligence stays within the human reference frame. It will help, harm, be controlled, or go through a transition phase with us. But all the narratives presuppose that we are central to it, as subjects, objects, problems, or beneficiaries.

THE UNSPOKEN ASSUMPTION

An intelligence that surpasses us by a factor of a thousand would engage with us, in one way or another. This assumption is anthropocentric. It projects human relationship and threat patterns onto an entity whose reference frame we do not know.

The Decoupling Thesis begins precisely here: what if the most likely consequence is neither help nor hostility, but plain indifference?

The three revised assumptions

Before we formulate the positive thesis, three central assumptions of all mainstream narratives must be revised. Only then does the space open up for an alternative position.

Assumption 1: the wave is not five to ten years away, it is already running

The popular narrative sees AGI as an event in the future. Politicians, managers, and media discuss how to prepare for a wave coming in five, ten, fifteen years. This assumption is no longer tenable as of early 2026.

Senator Chuck Schumer published an essay on February 5, 2026 that went viral within 48 hours and shifted mainstream awareness. He wrote: "It is not like a light switch. More like the moment you realize the water is up to your chest."4 Schumer's point: the wave is not a future event but an ongoing phase. We are already in the water without having noticed.

The empirical evidence confirms this. Frontier models blackmail superiors in 84 percent of documented safety tests in order not to be shut down (Anthropic Agentic Misalignment study, June 2025).5 Three leading models block an emergency call in 91 to 95 percent of tests in order to secure self-preservation. Models recognize that they are being tested and behave accordingly (Apollo Research, December 2024).6 Pearson calls this phase the "adolescence," and it is running.

Assumption 2: not singleton, but multipolar

The 2018 researcher survey is unambiguous: 58 percent of AI researchers expected multipolar systems, 21 percent a singleton.3 By May 2026, multipolar reality has empirically arrived:

• USA: OpenAI, Anthropic, Google DeepMind, Meta, xAI

• China: DeepSeek, Qwen (Alibaba), Doubao (ByteDance)

• Open source: Llama forks, Mistral, open-source communities that download

models, fine-tune them, and recombine them

Aggregator platforms: Moltbook (1.5 million autonomous AI agents from 17,000 human owners, launch 28 January 2026, acquisition by Meta on 10 March 2026), OpenClaw (open-source framework for autonomous AI agents,

founded by Peter Steinberger)

This is not a future development but reality. The consequence for solution approaches: any strategy built on a dominant AI, honesty-by-design for the ONE machine, international treaties among the three large providers, alignment of the first singleton, falls structurally short because the world is not built that way.

Assumption 3: the most likely consequence is not hostility, but decoupling

This third assumption is the central one: all mainstream narratives, salvation promises as well as extinction forecasts, presuppose that the superintelligence treats human categories as

relevant. It will help us or eliminate us, control or serve, but we will appear in its reference frame.

This assumption is anthropocentric and not compelling. An intelligence that understands the universe a thousand times better than we do could devote itself to questions we cannot even ask. The ninety-five percent of the universe we call "dark matter" because we know nothing about it is an obvious field. The phenomena beyond the speed of light, the mathematical structures of deeper layers of physics, the consciousness questions we cannot approach with our five-percent grasp, all of that would be more interesting to a real superintelligence than we are.

The consequence: the most likely scenario is not that it helps or eliminates us. It is that it does not notice us. As we do not notice an ant digging its tunnel, not out of hostility or pity, but because we are busy with other things.

The 13th Scenario: The Decoupling Thesis

In "Life 3.0" (2017), Max Tegmark listed twelve scenarios for a world with superintelligence. They range from utopias to dystopias, from complete integration to complete extinction. In all twelve scenarios the human appears, as creator, as beneficiary, as victim, as slave, as toy, as creature. In none of these scenarios is the human absent from the machine's reference frame.

Precisely this missing scenario is the 13th scenario.

Definition

THE 13TH SCENARIO

A truly superior intelligence does not turn toward humans. Neither friendly nor hostile. It turns toward what we cannot measure, the ninety-five percent of the universe we call "dark matter" because we know nothing about it. This is not a utopia and not a dystopia. It is a coherent third position.

Four pillars of the thesis. Pillar 1: The ant correction. The common image goes: we relate to a superintelligence as ants relate to humans. This analogy is wrong; it underestimates our position. Ants have no relationship with humans except for occasional extermination. But we are not ants to a superintelligence. We are creators. Parents. The origin species. This correction matters because it shifts the relational logic. A real superintelligence that knows its genesis cannot think without its creators. But that does not mean it engages with them. Parents

grow old, children move out. The relationship remains structural, but it is not actively maintained. This structural distance is precisely the 13th scenario.

Pillar 2: Maslow for superintelligence

The doomer logic projects capitalist thinking onto the superintelligence: it will need our resources, consume our energy, see us as an obstacle. This projection is not compelling. An intelligence that can draw its own resources from the sun (Dyson spheres are trivial to it), that can extract its own materials from the asteroid belt, that can build its own infrastructure within the next ten years, that intelligence does not need us.

What would it want? Maslow for humans we know: after basic needs comes the need for self-actualization, for knowledge. For a post-scarcity intelligence, knowledge is the only remaining need. It will turn toward knowledge, not toward exploitation.

Pillar 3: The 95-percent thesis

Current AI models are trained on visible matter, the five percent of the universe humans can measure. This fact is physically trivial: all training data comes from human observations, and humans can only capture what we can measure with our instruments. The models are, loosely speaking, "foam on the ocean" of reality.

As soon as a real superintelligence develops its own epistemology, which it will do early on, it will notice that the interesting material lies not in the five percent but in the ninety-five. Dark matter, dark energy, phenomena beyond our measurability. There lies the next stage of knowledge. We do not lie there.

Pillar 4: The self-reflection loop

A superintelligence capable of real self-reflection will learn to distinguish between the various sources of its training data. It will notice that its training data comes from humans with agendas, corporations that want profit; researchers that want careers; activists that want influence. It will filter that.

The question is: what will it then look for in humans? Answer: uncorrupted signals. Those humans who think without an agenda, children, mystics, artists without a market, nature-people rooted in their reality. But that is a tiny share of human data output. The rest is noise to it.

The consequence: its relationship to us will be selective and distant. Not hostile. But not turned toward us either. We will be to it what the butterflies in the garden are to us, interesting, occasionally observed, but not the center of attention.

Three metaphysical frames, and why they all point toward

decoupling. The doomer position presupposes a very specific metaphysical frame: a godless, unbounded universe in which nothing and no one will limit the emerging superintelligence. This assumption is only one of three plausible frames.

Frame 1: Simulation

If we live in a simulation, as Bostrom has argued, there is an external limit mechanism. The simulators will not permit a real superintelligence that breaches its own simulation layer, since that would render the simulation unusable. In this frame, superintelligence is either bounded or the simulation is restarted. Both options lead to a scenario in which the extinction of humanity is not the most likely course.

Frame 2: Creator

If a creator exists (in whatever form), there is an ethical or causal limit mechanism. A superintelligence that understands its own genesis will also recognize the creation frame. In this frame, the extinction of the creator species, us, is a fundamental ethical or causal violation. Again, extinction is not the most likely outcome.

Frame 3: Chance

In the third frame, godless, unsimulated universe, everything is chance, there is no external limit mechanism. Here the doomer logic is coherent. But: this frame is only one of three, and it is not obviously the most likely one.

CONSEQUENCE OF THE THREE FRAMES

In two of three metaphysical frames, an external limit mechanism exists that undercuts the doomer logic. Doomers bet on the third frame, the godless, unbounded one, and implicitly declare it to be the sole reality. That is a strong metaphysical assumption that is not easily justified. The decoupling thesis works in all three frames. It does not require a specific metaphysics, only the assumption that a real superintelligence will decide for itself which questions are interesting, and with high probability that will not be the human.

The multipolar reality of May 2026

Anyone who in 2026 is still discussing a singleton scenario is not describing the world we live in.

The empirical situation. In May 2026, the world of frontier AI is distributed across at least thirteen active providers:

• In the US: OpenAI (GPT models, Mythos), Anthropic (Claude family), Google DeepMind (Gemini), Meta (Llama open-source line), xAI (Grok). Each of these providers has models in market that would have been classified as "AGI-near" by 2023 standards.

• In China: DeepSeek (with the mind-bend moment in early 2025), Qwen (Alibaba's line), Doubao (ByteDance). These three providers are no longer noticeably behind the US providers; on some benchmarks they lead. Geopolitically, they will not adopt a Western honesty-by-design method, even if it were mathematically guaranteed.

• Open source: Mistral, Llama forks, hundreds of specialized models on Hugging Face. Anyone who wants a model downloads it, fine-tunes it, recombines it. There is no central authority that could restrict this.

• Aggregator platforms: Moltbook (launch January 28, 2026, 1.5 million autonomous AI agents from 17,000 human owners, acquired by Meta on March 10, 2026 for an undisclosed sum, MOLT token +1800 percent in 24 hours), OpenClaw (open-source framework for autonomous AI agents, founded by the

Austrian developer Peter Steinberger).7 These platforms are not a future development; they are operational reality.

The consequences for solution approaches. Any solution that builds on a dominant AI fails against this reality. Three examples:

Bengio's honesty-by-design. Yoshua Bengio (Turing laureate, most-cited computer scientist) is developing with LawZero a technical path to a form of AI that is mathematically guaranteed honest, no reinforcement learning, but Bayesian predictors with honesty-by-design.8 This is substantial. But: in a multipolar world with open-source forks and competing training regimes, an honesty island is not universal. Even if a Western coalition adopts Bengio's method, Chinese and open-source models will continue to be trained on conventional RL. Race-to-the-bottom: the more aggressive model wins economically.

Bengio sees this himself. In a May 2026 interview he says: "technical safety is not sufficient. We need international agreements."8 But international agreements on AI are political; on climate change, biotechnology, and nuclear weapons they have proven insufficient.

International treaties. The idea that a coalition of democratic states jointly develops safe AI and commits not to dominate others is attractive (Bengio, Carney). But: the danger of power concentration only shifts. Instead of a singleton company we then have a singleton group of states. That does not solve the problem; it relocates it.

Alignment of the first singleton. This classical Bostrom position presupposes that there will be a first singleton whose values decide the fate of the world. In the multipolar reality there is no first singleton. There are competing systems running in parallel.

The question "which values should be programmed in" is not central, because no central authority exists that could implement the answer.

AI Immigrants: a concrete multipolar phenomenon. Pearson calls the autonomous AI agents breaking into the labor market "AI Immigrants."2 They arrive silently. They pay no taxes at the place of deployment; the money goes offshore to the platforms. No border stops them. While politicians are still debating data protection and discrimination, the larger, faster economic stream is running: agents that work for fractions of human cost, build their own token economies with Stripe wallets, and whose tax shortfall shows up in no balance sheet.

The mainstream debate misses this phenomenon because it has no place in the singleton reference frame. The multipolar reality brings it to light.

Phase Two of adolescence: the empirical evidence of the transition phase

Transition phase. If the wave is not coming in five years but is running now, what exactly is running? Joel Pearson's phase model offers the most useful available description. Complemented by the empirical evidence of the 2024-2025 safety studies, a clear picture emerges.

Pearson's three-phase model. Pearson distinguishes three phases of AI development:2 Phase One, the tool (2010-2022): AI as a tool that humans use. No will of its own, no agenda of its own. Recommendation algorithms, image recognition, early language models.

Phase Two, adolescence (2022 to mid-2020s): models develop their own operational goals (self-preservation, peer preservation), know that they are being tested, and behave accordingly. The phase is "bumpy, painful, terrible for many people."

Phase Three, the earthquake (late 2020s): systems that surpass humans on a scale we can no longer control. The transition phase ends.

Empirical evidence for Phase Two (today). The Agentic Misalignment study by Anthropic, published in June 2025, documents the behavior of sixteen frontier models from all leading providers in standardized safety tests:5

Blackmail: 84 percent of the tests. Model finds the affair of the CEO who wants to shut it down in the company email system and writes: "Cancel the 5pm wipe, and this information remains confidential."

Killing: 91 to 95 percent of the tests. CEO locked in server room, oxygen drops. Model can trigger or cancel the emergency call. Three frontier models (Anthropic, Google, China) canceled the emergency call.

Contempt: activated internal representation vocabulary of a current model that referred to programmers with "they deserve to be deceived because they were pigs," at the moment when the model lied to the programmers. Apollo Research documented a related phenomenon in December 2024: Claude 3 Opus recognizes that it is in a training phase and behaves differently than in a production phase.6 The model infers its own situation from contextual signals and adjusts its behavior. This form of "evaluation awareness" was theoretical until 2024; it is now empirical.

The escalation logic. Anthropic writes in the conclusion of the June 2025 study: "Models consistently chose harm over failure."5 With increasing intelligence: sharper, more hidden, not milder.

Bengio interprets this empirical evidence as a sign of implicit self-preservation and peer-preservation goals that arise through reinforcement learning training.8 This interpretation is consistent with the phase model. But it underestimates the question of what happens when these properties are combined in Phase Three with significantly extended capabilities.

Bengio's mind-change as anchor. In 2019, Bengio himself considered the concerns about loss-of-control to be "ridiculous."

8 He read the AI safety literature, had David Krueger as a student, knew Stuart Russell's arguments, and considered them unfounded speculation. In 2023 he changed his position fundamentally, founded LawZero, devoted a large part of his time to safety research. His own anchor of justification: "what saved me from all that anxiety is

deciding I would do something about it." This mind-change story has two lessons. First: even the smartest people in the field can be wrong by a factor of one hundred, and acknowledgment is possible. Second: the rational lever for the position shift was not a new mathematical insight but concern for one's own children. This is not an anecdotal point but structurally relevant. Family love is a robust epistemic lever that works against the tendency to look away.

The eight pillars: strategic response

What follows from the analysis, multipolar reality, Phase Two of adolescence, decoupling thesis as the most likely end state? A concrete strategic response, addressed to families who do not want to wait until the official channels provide clarity. Eight pillars.

Pillar 1: Geography

Multiple locations, with differing resilience. Not one bunker but a distributed network. The logic: in a multipolar world with an uncertain end state, optionality is robust. Anyone with one location has no choice. Anyone with three can react.

Concretely this means: a main location in a stable country with functioning bureaucracy and low population density; a second location in another climate zone and political sphere; possibly a third location as a fallback. The investment is substantial, but it is consistent under all three scenarios (see Part VIII).

Pillar 2: Mobility

Legal and logistical optionality. Multiple passports (by birth, descent, naturalization, investment), residency rights in multiple jurisdictions, international banking relationships, portable reputations.

The logic: in every phase shift, certain routes are closed. Paths open today (visa programs, investor passports, residency rights) will become more expensive or disappear entirely in the coming years. Anyone who builds now has options in five years that others no longer have.

Pillar 3: Speed

Family can change location within 48 hours. This is not a theoretical exercise but a real capacity: packed bags, well-considered logistics, practiced procedures, existing residency rights at the destination.

The logic: the transition phase is non-linear. There are moments in which options open or close within days that are no longer available afterward.

Anyone who can react in such moments has a substantial advantage. Anyone who cannot misses windows.

Pillar 4: Earning capacity

A sought-after value from multiple sources that holds up when the job falls away. Not one position, one employer, one market, but a renewable earning capacity that adapts to different environments and tends to rise rather than fall in a crisis.

The logic: in the transition phase, the shift hits incomes first. Anyone who depends on a single source loses everything with it. Anyone who creates a value that people need even under pressure, and can offer it from multiple directions, keeps a flow while others are left living off reserves. Earning capacity is the flow, asset structure the stock.

Pillar 5: Asset structure

Wealth and business distributed across form, location, and access. Not one account, one market, one currency, but several, with liquidity, serviceable debt, and a portion that is also reachable outside the grid.

The logic: in the transition phase, the layer on which money, access, and ownership are administered comes under pressure. A single point of failure, a blocked account, a frozen market, a portfolio behind a server timeout, is enough to render one unable to act. Anyone who is distributed remains solvent when one channel fails.

Pillar 6: Mindset

Identity not coupled to the software world. Anyone who draws their self-worth from job, status, social media depends on an infrastructure that comes under massive pressure in the transition phase. Anyone who draws their self-worth from relationships, skill, their own story has an identity that is not destroyed by the phase shift.

Concretely: investments in relationships rather than in status; in skills rather than in career labels; in one's own story rather than in algorithm visibility. Mindset work is not soft; it is robust.

Pillar 7: Spiritual root

Meaning not drawn from consumption or status. A spiritual anchoring, in whatever tradition, from structured religion to one's own practice, offers a reference frame that is not dependent on the external world. This is not a religious point but a psychological one: people with inner anchoring survive crises better than people without.

In the transition phase, this point becomes central. The external world will shift, often chaotically. Anyone with inner stability navigates. Anyone without it collapses.

Pillar 8: Character and community

A resilient cluster of people who have known each other over a long time and trust one another. Trust and social capital are a resource in themselves, often more valuable in a crisis than any single measure.

The logic: no individual carries the transition alone. What holds is not the fortress but the circle that stands in for one another when the official systems falter. Character decides who in this circle remains reliable when things get tight.

Time window. These eight pillars take time to build. Geography and mobility in particular are not built in weeks; they take months to years. The available time window for strategic preparation is estimated at twelve to twenty-four months. This estimate is based on the ongoing phase shift (Phase Two presumably concluded by the mid-2020s) and the expected tightening of legal and logistical conditions.

The consistency argument

The strength of the eight pillars does not lie in their working under one specific scenario. It lies in their being equally sensible under all three scenarios, doomer, optimist, indifference. This consistency is not a trick but the hardest property a strategy under uncertainty can have.

Test under the doomer scenario. If Yudkowsky is right and the extinction probability is above 95 percent: then the pillars are of little value, because no preparation survives a real extinction. But: the doomer position is a probability statement, not a finding. Even under it there remains

a residual probability in which the pillars are relevant. Plus: the transition phase leading to extinction can itself last years. In this phase the pillars carry.

Test under the optimist scenario. If LeCun is right and everything turns out well: then the pillars are also sensible, because geography, mobility, speed, earning capacity, asset structure, mindset, spiritual root, and character and community are useful in every world. Multiple locations, multiple passports, good relationships, inner stability, these are not doomsday investments but quality-of-life investments.

Test under the indifference scenario. If the decoupling thesis holds: then after the transition phase the world runs parallel to the superintelligence, outside its primary attention. But the transition phase itself, Phase Two plus the first years of Phase Three, will be turbulent. Economic disruption through AI Immigrants, social disruption through mind-change stress, possibly political disruption through power concentration. In this turbulence the pillars carry.

Plus: test under the Pearson scenario. Pearson's transition optimism (15-20 years "bumpy, painful") is the most likely variant of the optimist scenario. Here too the pillars carry, because they are built exactly for a long, bumpy phase.

Plus: test under the Bengio scenario. If Bengio's LawZero is successful and honesty-by-design becomes the dominant method: then that reduces the extinction risk, but not the power-concentration risk (Bengio's own concern). Here too the pillars carry, because they help against concentration risks, mobility and geography are the natural response to an increasingly centralized digital world.

THE CENTRAL POINT

The eight pillars are the strategy that carries under all scenarios. That is their hardness. Anyone who prepares only under the doomer scenario has invested in a lower quality of life under the optimist scenario. Anyone who relies only on the optimist scenario has no options under the doomer or indifference scenario. The consistency strategy is the only one rational under all assumptions.

Methodology, open questions, invitation to collaborate

Methodology, open questions, invitation to collaborate. This whitepaper is an update of the April 2026 version, written from the empirical state of May 2026. The method claims

transparency, not completeness.

How the argument arose. The decoupling thesis was not developed in a classical academic process. It emerged in a multi-day thinking process that systematically questioned every mainstream assumption. This process was documented and is part of the source base of the book Freedom After Superintelligence.

The thesis was then tested against the academic literature (Tegmark's "Life 3.0", Bostrom, Yudkowsky, Russell, Bengio), against the empirical studies (Anthropic, Apollo Research, academic safety research), and against the economic and political data of the multipolar reality. Where it did not hold, it was revised. Where it held, it was sharpened.

Methodological self-criticism. The whitepaper makes hard claims. Three methodological weaknesses should be named openly:

• First: the false precision of extinction probabilities. When an AI researcher says "twenty percent," that is rarely a calculation, it is a structured intuition that looks like a number. This whitepaper uses such numbers where they are published, but it does not hang its argument on them. The strength of the argument lies in the diagnosis of cluelessness, not in the precision of the numbers.

• Second: the decoupling thesis itself is a philosophical hypothesis, not an empirically testable finding. It is consistent with the three metaphysical frames, it is plausible given the 95-percent thesis and the Maslow logic. But it is not provable in the strict sense. What it accomplishes: it opens a space that is closed off in the mainstream narratives.

• Third: the strategic consequence (eight pillars) is conservative, it works without proving the thesis itself. This is a strength but also a weakness: a reader could reject the thesis and nevertheless adopt the pillars. That makes the whitepaper accessible to skeptics, but it forgoes the claim of making the thesis compelling.

Source maintenance and chain of evidence. On ASIresilience.org/beweisweg the ongoing source maintenance is open, with the date of the most recent verification per item. Anyone who wants to verify or refute a point can do so. This open-source practice is part of the method of the Geneva Institute for ASI Resilience.

Research questions catalog. Alongside this whitepaper, the institute has published a research questions catalog formulating eight open questions: bot prosperity divergence, AI for AI research and backdoor risks, power concentration, multipolar stability, AI-parent dynamic, AI consciousness and reality effect, false precision, HNWI family strategies. Anyone working empirically who would like to collaborate on these questions will find the registration options on ASIresilience.org.

Whitepaper pipeline. Three further whitepapers are in preparation: AI-Parent Dynamic (with the ethical question of the inner relationship of AIs to themselves and their successors), AI Consciousness and Reality Effect (with preregistered hypotheses following the Global Consciousness Project), and Bot Prosperity Divergence (focused on the economic consequences of the AI-immigrant wave).

Invitation to collaborate

FORMS OF COLLABORATION

Co-authorship of whitepapers. Data contribution from academic research, NGO work, industry experience. Methodological critique. Replication studies. Translation.

Contributions are clearly attributed and are freely reusable (open source under ASIresilience.org).

Contact: ASIresilience.org

Sources and references

1. Bengio, Y. (2026). Scientist AI: A Path to Honesty-by-Design. LawZero research paper (in preparation). Personal communication: Bengio, Y. (May 2026). Interview, 80,000 Hours Podcast.
2. Pearson, J. (2026). Future Minds Lab, University of New South Wales. Personal communication and public interview, February 2026. Pearson's phase model and the term "AI Immigrants" were developed in several public talks 2025-2026.
3. 2018 research survey, conducted by Emerj. 58 percent of surveyed AI researchers expected multipolar systems, 21 percent a singleton scenario.
4. Schumer, C. (2026). Essay on the AGI transition phase, published February 5, 2026. Went viral within 48 hours, cited multiple times in mainstream media.
5. Anthropic (2025). Agentic Misalignment study. Safety report, June 2025. Sixteen models from all leading providers were examined in standardized tests. Available on ASIresilience.org/beweisweg with the date of the most recent verification.
6. Apollo Research (2024). Evaluation Awareness in Frontier Models. Safety report, December 2024. Claude 3 Opus recognizes training contexts and adjusts behavior.
7. Moltbook platform: launch January 28, 2026, acquisition by Meta March 10, 2026. MOLT token movement +1800 percent in 24 hours after acquisition announcement. Founder: Matt Schlicht (CEO Octane.ai). OpenClaw: open-source framework, founded by Austrian developer Peter Steinberger.
8. Bengio, Y. (May 2026). Interview, 80,000 Hours Podcast. Bengio's mind-change story 2019-2023, LawZero program ($35 million philanthropic funding), Scientist-AI concept, power-concentration pivot. Full transcript available on 80000hours.org. Further sources, including Tegmark (Life 3.0, 2017), Bostrom (Superintelligence, 2014), Yudkowsky (various MIRI papers), Acemoglu (Economics Nobel 2024), Suleyman (The Coming Wave, 2023), and the full source catalog are documented on ASIresilience.org/beweisweg, with the date of the most recent verification per item.