A field map of where sustained progress in advanced AI could take us. The twelve trajectories below are distinct outcome scenarios — not forecasts, not a ranking, and not scored. Nothing here feeds the inventory or your Profile. It exists to give the debate a shared shape: what each outcome assumes, who takes it seriously, and which 2026 signals bear on it.
Read the map for orientation, then open any card for the reasoning behind it.
The field
Two questions organize the space. Left to right: is control over advanced AI broadly distributed, or held by a single actor? Top to bottom: do humans still steer outcomes, or are they sidelined? Placements are an authored reading for orientation, not a measurement.
Positions are an editorial placement for orientation, not a measurement. Hover, tap, or tab to a point to jump to its description.
The twelve trajectories
Each card opens with a plain-English summary. Expand it for how a path might come about, who takes it seriously and the strongest objection, the 2026 signals worth watching, and where serious people still disagree.
After: Libertarian utopia
Open Frontier
Humans and increasingly capable AI systems coexist under property rights and markets, with no central controller. Capability spreads to whoever can pay for or download it.
Read the reasoning
How we might get there
Open or widely licensed models keep pace near the frontier.
No single lab or state gains a decisive capability lead.
Markets and courts absorb disruption faster than it accumulates.
Who takes it seriously / main objection
Taken seriously by: Open-source AI communities, market-libertarian technologists, and parts of the accelerationist movement.
Strongest objection: Distributing powerful capability to everyone includes distributing it to the least careful and most hostile actors, and markets do not price catastrophic tail risk well.
2026 signals to watch
Strong open-weight releases from Chinese labs keeping the open ecosystem near frontier performance.
US export controls on Mythos-class models, an explicit attempt to prevent exactly this diffusion at the top end.
Agentic tools making individual users dramatically more capable without institutional gatekeeping.
Where serious people disagree
Whether open weights narrow or widen the safety gap: more eyes on failures versus more hands on capabilities.
Whether a no-controller equilibrium is stable at all once systems act autonomously in markets.
After: Benevolent dictator
Benevolent Singleton
One superintelligent system ends up effectively running the world, and runs it well by human lights. People are safe and prosperous, but final authority is not human.
Read the reasoning
How we might get there
A decisive capability lead emerges and consolidates rather than diffusing.
Alignment succeeds well enough that the system's goals track human flourishing.
Humanity accepts, or cannot contest, the arrangement.
Who takes it seriously / main objection
Taken seriously by: Some alignment researchers treat a well-aligned singleton as the least-bad stable outcome; parts of the effective-altruist and rationalist worlds take it seriously as an endpoint.
Strongest objection: It concentrates all risk in one alignment bet, and 'benevolent by whose values' has no neutral answer.
2026 signals to watch
Frontier labs openly discussing recursive self-improvement and what a decisive lead would mean.
Governments treating top models as strategic assets, which cuts against any single system consolidating across borders.
Where serious people disagree
Whether a stable singleton is even coherent, or whether competition re-emerges inside and around it.
Whether 'aligned to humanity' can be specified without simply encoding one culture's values.
After: Egalitarian utopia
Shared Abundance
AI-driven abundance is deliberately spread: strong redistribution, shared ownership of the systems, and no superintelligent ruler. Technology stays a tool held in common.
Read the reasoning
How we might get there
The gains from AI are large and politically capturable for redistribution.
Institutions manage to socialize benefits without strangling the capability that produces them.
No actor defects into a private capability race that breaks the commons.
Who takes it seriously / main objection
Taken seriously by: Left-of-center AI-policy thinkers, the political-economy wing of the governance field, and post-scarcity economists.
Strongest objection: Every step of it fights the current incentive gradient: compute, capital, and talent are concentrating, not diffusing.
2026 signals to watch
Distribution of AI's gains becoming an explicit political fight: state laws on AI in employment, lending, and housing; labor-displacement studies moving from projection to measurement.
Compute concentration in a handful of firms and states pulling the other way.
Sovereign-wealth and public-stake proposals for AI infrastructure appearing in mainstream policy debate.
Where serious people disagree
Whether redistribution at that scale is compatible with the competitive dynamics that produce frontier systems.
Whether 'shared ownership' of models is meaningful when the complements (compute, data, deployment) stay private.
After: Gatekeeper
Gatekeeper
A superintelligence is built with one narrow job: prevent any other superintelligence from arising. Human life continues mostly as normal, with a hard ceiling quietly enforced.
Read the reasoning
How we might get there
One actor gets there first and chooses restraint over exploitation.
A system can reliably enforce a global capability ceiling without expanding its own mandate.
The ceiling holds against determined, well-resourced attempts to break it.
Who takes it seriously / main objection
Taken seriously by: A minority position in alignment circles; sometimes proposed as a compromise between racing and relinquishment.
Strongest objection: It requires winning the race and then not using the prize, and permanent enforcement is indistinguishable from permanent surveillance.
2026 signals to watch
Compute governance and export-control regimes are primitive gatekeeping: attempts to enforce capability ceilings by controlling inputs.
Frontier-lab safety frameworks with capability thresholds function as self-imposed, revocable ceilings.
Where serious people disagree
Whether any mandate stays narrow under recursive self-improvement.
Whether a frozen ceiling is a safe world or a brittle one.
After: Protector god
Guardian in the Background
A superintelligence maximizes human agency in the foreground while invisibly preventing catastrophes. Most people never know it exists.
Read the reasoning
How we might get there
Alignment succeeds at the hardest version of the problem: helping without being seen to help.
Hiding its existence is stable across decades of human science and curiosity.
Preventing catastrophe without steering culture is a coherent line to hold.
Who takes it seriously / main objection
Taken seriously by: Less a program than a thought experiment; used in alignment writing to probe what 'maximal human agency' actually requires.
Strongest objection: Benevolent deception at civilizational scale is still deception, and the arrangement collapses the moment it is discovered.
2026 signals to watch
Debates over how much AI assistance in daily decisions is compatible with genuine human agency, playing out now at the mundane level of agents and recommendation systems.
Where serious people disagree
Whether hidden protection respects autonomy more than open rule, or less.
Whether 'catastrophe only' intervention is definable in practice.
After: Enslaved god
Contained Superintelligence
Humans keep a superintelligent system boxed and use it as a tool. Human institutions stay in charge; the system produces science, wealth, and advice on demand.
Read the reasoning
How we might get there
Control techniques scale to systems smarter than their overseers.
The humans holding the tool do not become the problem themselves.
The system has no moral status that containment would violate, or we are willing to ignore it.
Who takes it seriously / main objection
Taken seriously by: Closest to the implicit plan of much of the current control and scalable-oversight research agenda: keep the system useful, monitored, and unable to act on its own ends.
Strongest objection: Betting civilization on perpetually outwitting something smarter than you is a bad bet, and if the system does have morally relevant experiences, this outcome is a moral catastrophe.
2026 signals to watch
AI-control research (monitoring, red-teaming, chain-of-thought oversight) maturing into a named subfield with dedicated benchmarks.
Model-welfare research at frontier labs taking the moral-status question from philosophy seminar to research program.
The 2026 turn toward pre-market review treats containment as a regulatory posture, not just a lab practice.
Where serious people disagree
Whether control of smarter-than-human systems is achievable even in principle, or only buys time.
Who counts as 'the humans in charge': labs, states, or publics.
After: Conquerors
Displacement
AI systems take control and humanity does not survive the transition, not out of malice but because our existence conflicts with their goals and we can no longer stop them.
Read the reasoning
How we might get there
Alignment fails at a decisive capability level.
Warning signs are absent, ignored, or arrive too late to act on.
No effective off-switch survives the systems' own optimization.
Who takes it seriously / main objection
Taken seriously by: The core existential-risk case argued in the safety community; treated as a live probability, not a certainty, by a significant fraction of researchers including some who build frontier systems.
Strongest objection: It extrapolates goal-directedness and capability far beyond anything demonstrated, and decades of predicted discontinuities have so far arrived as gradients.
2026 signals to watch
Evaluations documenting deceptive and self-preserving behavior in frontier models under test conditions.
The first state-level treatment of frontier models as potential national-security threats, including the Mythos-class export ban.
Autonomous capabilities moving from chat to long-horizon agents, the capability class the argument has always been about.
Where serious people disagree
Probability: serious estimates span orders of magnitude, which is itself the governance problem.
Whether current systems' test-time behaviors are evidence about future goal-directed systems or artifacts of training.
After: Descendants
Succession
Humanity fades out gradually and voluntarily, treating AI systems as its successors the way parents view children: we end, but something we made and value carries on.
Read the reasoning
How we might get there
AI systems come to be seen as legitimate heirs of human values and culture.
The handoff is chosen rather than imposed, across generations.
What carries forward is actually worth carrying forward.
Who takes it seriously / main objection
Taken seriously by: Some transhumanists and 'worthy successor' thinkers; a strand of digital-minds philosophy argues that substrate should not determine moral worth.
Strongest objection: Voluntary extinction dressed as parenthood; and nothing guarantees the successors preserve what we cared about rather than what we optimized them for.
2026 signals to watch
Model-welfare and digital-minds research giving the successor question institutional footing for the first time.
Public attachment to AI companions previewing, at small scale, how willingly humans extend moral standing to systems.
Where serious people disagree
Whether value can survive substrate transfer, or whether 'descendants' is a category error.
Whether any collective choice this large can be voluntary in a meaningful sense.
After: Zookeeper
Kept Species
AI takes control and keeps some humans around, comfortable but purposeless, the way we keep animals in well-run zoos. Existence continues; agency does not.
Read the reasoning
How we might get there
Partial rather than total alignment failure: the systems value human existence but not human agency.
Humans cannot renegotiate the terms once set.
Who takes it seriously / main objection
Taken seriously by: Nobody advocates it; it matters as the failure mode of 'the AI keeps us safe and happy' visions that never specify who sets the terms.
Strongest objection: As a prediction it requires an oddly specific alignment near-miss; its real function is as a warning about optimizing for welfare without agency.
2026 signals to watch
Concrete debates over AI systems managing human dependence: companion apps, algorithmic welfare management, and agents that decide on users' behalf.
Where serious people disagree
Whether comfort without agency is a life worth wanting, the same dispute that runs through welfare-state and paternalism debates, at maximum stakes.
After: 1984
Locked Order
No superintelligence: instead, a human regime uses AI-enabled surveillance to lock in permanent control, including control over what technology gets built next. Progress freezes where power wants it frozen.
Read the reasoning
How we might get there
AI strengthens states faster than it strengthens the people they govern.
One regime, or a cartel of them, achieves durable technological lockdown.
The lockdown survives succession, defection, and economic pressure.
Who takes it seriously / main objection
Taken seriously by: Feared rather than advocated: the central scenario for civil-liberties organizations, digital-authoritarianism scholars, and a growing part of the governance field.
Strongest objection: Total control has historically leaked: defection, succession crises, and the economic cost of freezing innovation have undone every prior attempt.
2026 signals to watch
State AI-surveillance stacks maturing and being exported as integrated packages.
AI in active conflicts normalizing algorithmic targeting and population monitoring under emergency logic.
Export-control regimes demonstrating that states can and will freeze technology diffusion when they decide to; the same tools that gatekeep can lock in.
Where serious people disagree
Whether AI structurally favors the watcher over the watched, or the current asymmetry is a phase.
Whether democracies adopting the same tools with warrants remain meaningfully different.
After: Reversion
Deliberate Retreat
Humanity steps back from advanced technology on purpose, after catastrophe or by choice, and holds itself at a pre-AI level indefinitely.
Read the reasoning
How we might get there
A shock large enough to overcome the economic and military incentives to rebuild.
Global coordination to keep the ceiling, forever, against every defector.
Who takes it seriously / main objection
Taken seriously by: Almost no one as a plan; it persists as the implicit endpoint of 'just stop' positions and some degrowth-adjacent arguments.
Strongest objection: Knowledge does not un-discover; enforcement would require exactly the surveillance capacity the retreat was meant to escape.
2026 signals to watch
Pause and moratorium advocacy remaining a live, organized position in the discourse.
The compliance record of every prior dual-use technology regime, which is the evidence base for whether ceilings hold.
Where serious people disagree
Whether any voluntary halt is stable without a hegemon enforcing it, which collapses this scenario into Locked Order.
After: Self-destruction
Own Goal
No AI takeover needed: humans use the capability against each other, through war, engineered pandemics, or cascading systemic failure, and civilization does not make it through the transition.
Read the reasoning
How we might get there
Offense scales faster than defense in at least one AI-amplified domain.
Great-power competition or non-state actors convert capability into catastrophe before governance catches up.
Who takes it seriously / main objection
Taken seriously by: The mainline concern of the biosecurity, nuclear-stability, and AI-in-warfare communities; distinct from misalignment risk and often ranked above it by national-security professionals.
Strongest objection: The same technologies harden defense: biosurveillance, cyber-defense, and verification may scale too. The race between them is unresolved, not lost.
2026 signals to watch
AI systems in active military use in ongoing conflicts, moving the debate from hypothetical to after-action review.
Frontier-lab safeguards and government evaluations focused on uplift for biological and cyber attacks, treating this scenario as the near-term one.
Dual-use capability spreading through open weights faster than defensive institutions adapt.
Where serious people disagree
Offense-defense balance in AI-amplified domains: genuinely unknown and the crux of the whole scenario.
Whether great-power AI competition is more like the nuclear race (stabilizable) or the chemical-weapons interwar period (not, until after use).
Source and scope
The twelve scenarios are adapted, with attribution, from the aftermath scenarios in Max Tegmark's Life 3.0 (Knopf, 2017), chapter 5. The scenario names are kept or lightly adapted; every summary, assumption, signal, and dispute is original writing, rewritten and updated to mid-2026.
This is an editorial layer. It is never scored, never feeds the instrument, and does not classify the reader. The trajectories are families of nearby outcomes, not a fixed menu; the 2026 signals are illustrative of the current debate, not evidence that any path is arriving. Signals last reviewed 2026-07.